Micro-loans, Insecticide-Treated Bednets and Malaria: Evidence …tarozzi/TarozziEtAl2013RCT.pdf · 2013-09-16 · Micro-loans, Insecticide-Treated Bednets and Malaria: Evidence from

Micro-loans, Insecticide-Treated Bednets and Malaria:

Evidence from a Randomized Controlled Trial in Orissa (India)

Alessandro Tarozzi, Aprajit Mahajan, Brian Blackburn

Dan Kopf, Lakshmi Krishnan, Joanne Yoong ∗

AbstractWe describe findings from the first large-scale cluster randomized controlled trial ina developing country that evaluates the uptake of a health-protecting technology,insecticide-treated bednets (ITNs), through micro-consumer loans, as compared tofree distribution and control conditions. Despite a relatively high price, 52% of samplehouseholds purchased ITNs, highlighting the role of liquidity constraints in explainingearlier low adoption rates. We find mixed evidence of improvements in malaria indices.We interpret the results and their implications within the debate about cost sharing,sustainability and liquidity constraints in public health initiatives in developing coun-tries.JEL: I1,I3.Key words: Malaria, Bednets, Microfinance, Public Health.

∗We are very grateful to a number of colleagues, collaborators and institutions for making this studypossible. In particular: BISWA and especially Khirod Chandra Malick, Dipti Pattnaik and Asish Sahoofor invaluable help in designing the micro-loan products and for facilitating access to study villages; AnupRoy, Benita Sarah Matthew, Deepak Nayak, Projjal Saha, Sudhansu Behera and all survey monitors foroutstanding project supervision and implementation; Dr. Madan Mohan Pradhan, at the National VectorBorne Diseases Control Programme, Orissa, for help at various stages of the interventions; Annie Duflo andthe Center for Micro Finance for invaluable help in making this study possible; Ravi and Saurabh Singhalof Biotech International for their generous donation of ITNs for the study and Binax for donating part ofthe materials necessary for measuring hemoglobin; Dr K S Sharma and his team at the Malaria ResearchCentre Field Station in Rourkela for their essential help in the validation study of the rapid diagnostictests; the Dwight Mount Division of Parasitic Diseases, Centers for Disease Control and Prevention, fortesting ITNs for insecticide levels; the Center for Micro Finance (Chennai, India), the Stanford Presiden-tial Fund for Innovation in International Studies, the Stanford Center for International Development, theStanford OTL Research Incentive Fund, RAND Corporation and the Duke Arts & Sciences Committee onFaculty Research for financial support; Jason Blevins, Patricia Foo and Kristin Johnson for outstandingresearch assistance; James Berry, Pascaline Dupas, Seema Jayachandran, Leigh Linden, Andrea Locatelli,Grant Miller, Wendy Prudhomme O’Meara, Elisa Sicuri, Duncan Thomas, Paul Wise, many participants toseminars and conferences and especially Marianne Bertrand (the Editor) and anonymous referees for con-structive comments and suggestions that substantially improved the manuscript. This study was partiallyfunded by Award Number R03AI078119 from the National Institute of Allergy and Infectious Diseases andby the Marie Curie International Incoming Fellowship FP7-PEOPLE-2011-IXF, Proposal ID 298904. Theauthors are solely responsible for the content of this paper. Tarozzi (corresponding author), Departmentof Economics and Business, Universitat Pompeu Fabra and Barcelona GSE, [email protected], Department of Economics, UCLA, [email protected]. Blackburn, Stanford University Schoolof Medicine, Division of Infectious Diseases and Geographic Medicine, [email protected]. Kopf,[email protected]. Krishnan, [email protected]. At the time of their involvement in the project, Kopfand Krishnan were employed by the Centre for Micro Finance (CMF), part of the Institute for FinancialManagement and Research (IFMR, Chennai, India). Yoong, National University of Singapore and RANDCorporation, joanne [email protected].

1 Introduction

A number of recent empirical studies have demonstrated that without free distribution or

substantial subsidies, the adoption of highly beneficial health-protecting technologies remains

very low among the poor in developing countries. Demand has been shown to be remarkably

price-elastic, with even small levels of cost-sharing leading to huge declines in adoption. In

Kenya, Kremer and Miguel (2007) found that a 20% co-pay for drugs to eliminate intestinal

worms reduced uptake from 75 to 19%. In urban Zambia, Ashraf et al. (2010) estimated a

price elasticity of −0.6 for the demand of a relatively inexpensive water disinfectant, effective

for the prevention of waterborne diseases that are especially dangerous to young children.

Kremer et al. (2009) documented only a 10% uptake when a similar product was offered at

half-price in Kenya. In rural Kenya, Cohen and Dupas (2010) found that a remarkable 90%

subsidy reduced adoption of insecticide treated bednets to 10%, relative to 99% achieved

with free distribution. In rural Zambia, subsidization did not increase bednet ownership

rates among the poorest households (Agha et al. 2007).

Liquidity constraints have been hypothesized to be a key reason for such low adoption

rates, because several health products require investing sums that may be non-negligible

for poor households (Dupas 2012a). Free provision or heavy subsidization are thus being

advocated by some quarters, especially in the presence of externalities in adoption such

as in the case of insecticide treated nets (WHO 2007) or de-worming drugs (Kremer and

Miguel 2007).1 However, budget constraints often impose serious limits on the ability of

public health campaigns to protect all those at risk. In addition, there has been much recent

debate on the sustainability of development initiatives, with advocates citing cost-recovery

as a crucial criterion for evaluating poverty reduction, health and education programs (Alam

and Ahmed 2010, Sarriot et al. 2011, Smith 2010). Cost-sharing may also help targeting

subsidies towards users with higher marginal benefits, although liquidity constraints will

limit such objectives if those at risk are also less able to pay.

When heavy subsidization is not possible but liquidity constraints are a key determinant

of low demand, micro-loans may offer a promising option in the search for sustainable public

health initiatives. This paper describes findings from the first large-scale cluster randomized

controlled trial (RCT) in a developing country context that evaluates the adoption and health

impacts of a health-protecting technology offered with micro-consumer loans, relative to free

distribution or control conditions. Specifically, we evaluate the effectiveness of micro-loans

1See e.g. Hammer (1997), Gersovitz and Hammer (2004, 2005) for an examination of economic approaches

towards health and infectious diseases in particular.

2

at increasing ownership and use of insecticide treated bednets (ITNs), and ultimately at

reducing the burden of malaria in areas of rural Orissa (India) where the disease is endemic.

Transmitted by Anopheles mosquitoes, malaria represents an enormous global health

burden, with a worldwide incidence of 300-660 million cases annually, 80 million in India

alone.2 One third of the human population is estimated to live in areas at risk for the most

severe form of malaria, caused by Plasmodium falciparum (Snow et al. 2005). The negative

association between the disease and economic growth and the accumulation of human capital

has been long recognized, although studies that convincingly document and quantify causal

links are relatively recent within the economics literature.3 Numerous randomized trials

have shown that with high coverage and/or high usage rates ITNs are efficacious at reducing

malaria-related morbidity and mortality, as documented in the extensive survey in Lengeler

(2004). However, ITN adoption in most malarious areas remains very low and public health

agencies frequently have insufficient resources to provide universal ITN coverage. In such a

context, a more sustainable approach focusing on cost-recovery may be desirable, but it may

lead to the exclusion of vulnerable individuals who do not have access to sufficient funds.

Our field experiment was conducted in 141 villages in rural Orissa, in collaboration with

BISWA (Bharat Integrated Social Welfare Agency), a micro-lender with a large presence

in the state. After a baseline household survey, completed in the spring of 2007, we ran-

domly assigned villages to three equally sized groups. A control group received no further

interventions, while lender clients in a second group received at no cost a number of ITNs

depending on household composition. Clients from the third group were offered contracts for

the purchase of ITNs and re-treatments, using consumer loans with a one-year repayment

period. The ITN offer price was not subsidized and included a mark-up to cover delivery

and overhead costs to BISWA. The price was not negligible, corresponding approximately

to 3-5 times the local daily agricultural wage.

This paper has two specific aims. First, we evaluated to what extent the offer of small

loans for purchasing ITNs led to increases in ownership, even among poor households. To

the best of our knowledge, this is the first large-scale cluster RCT to evaluate the efficacy of

a public health program where a health-protecting technology was provided at full cost but

allowing for repayment over time, as compared to both control conditions or free distribution.

Fink and Masiye (2012) describe the result of a later study in Zambia, where bednets were

2Snow et al. 2005, Korenromp 2005. For a comprehensive survey of the disease, including its epidemiology,

pathology and treatment see White (2009).3See Gallup and Sachs (2001), Sachs and Malaney (2002), Malaney et al. (2004), Hong (2007a), Hong

(2007b), Barreca (2010), Bleakley (2010), Cutler et al. (2010), Lucas (2010), Kitchens (2012).

3

offered on credit to farmers with access to agricultural loans. Devoto et al. (2012) study

adoption of piped water contracts offered on credit in urban Morocco, but focus on how

information and counseling affected loan applications open to all study subjects. We also

evaluate the cost-effectiveness of micro-loans when compared to free distribution, taking

into account that we observed partial repayment rates. In order to further gauge the role of

liquidity constraints as a barrier to demand, we also studied uptake of ITNs offered for cash,

although this intervention was conducted at a later time.

Second, we evaluated the impact of the alternative ITN delivery mechanisms on different

malaria indicators. Our data include results from thousands of blood tests that allow us to

estimate changes in malaria prevalence (the fraction of infected individuals at a given point

in time) as well as in hemoglobin levels. We also study changes in malaria incidence (the

number of illness episodes over a period of time), although these were respondent-reported.

The impact of ITNs distributed free of cost on malaria indices has been studied extensively,

although all but one of the 22 studies reviewed in Lengeler (2004) were ‘efficacy’ trials, that

is, conducted under highly controlled conditions generally leading to high coverage and use

rates. In contrast, our study could be seen as an ‘effectiveness’ trial, carried out on a large

scale and without potentially invasive surveillance of malaria indices and ITN usage. In

addition, our program did not seek universal community-level coverage but only targeted

BISWA-affiliated households. This led to low coverage, a condition that may be important

for public health interventions where externalities are important. Ours is also the first large-

scale RCT that analyzes the impact of ITNs on malaria indices in India.4

The rest of the paper is organized as follows. Section 2 describes the study area, the

RCT design and the data. Section 3 describes the impacts on ITN adoption and (self-

reported) usage and examines the role of liquidity constraints on demand for ITN. Section

4 discusses the impacts on malaria indices as measured both through blood tests and recall

data, after clarifying the features of each indicator and their inter-relationship. This section

also discusses the findings in light of the epidemiological and public health literature on the

impact of ITNs on malaria indices. Section 5 briefly considers cost-effectiveness of free ITN

provision versus sales on credit. Finally, Section 6 summarizes and interprets the results and

highlights limitations. Because of space constraints, we will refer the reader to the Appendix

for a number of details and additional results.

4A number of studies have been conducted in Orissa and elsewhere in India but they lack an appropriate

control group and/or have insufficient sample size, see Lengeler (2004, p. 16) for references.

4

2 Location, Study Design and Data

This study was carried out in 141 villages from five districts in rural Orissa, the most

highly malaria-endemic state in India (Kumar et al. 2007), and conducted in collaboration

with BISWA, a micro-lender with a large local presence. Study locations were selected

by stratified random sampling from a list of 878 villages with BISWA presence.5 A pre-

intervention baseline survey was completed in May-June 2007 for a random sample of 1,844

households. Within each village, 15 households were randomly selected from all those with

preexisting BISWA accounts as of November 2006, regardless of whether they had an active

loan at that time (all were selected if fewer than 15 were present in the BISWA rosters).

Two key malaria indicators (malaria prevalence and hemoglobin levels, Hb) were mea-

sured with rapid diagnostic tests (RDTs), and the results were immediately communicated to

the subjects. These tests require very small blood samples and deliver results within minutes

(see Appendix A.2 for details). Individuals targeted for blood tests included all pregnant

women, children under the age of five (U5) and their mothers, and one randomly selected

adult (age 15-60). The malaria RDT detects current or recent infections accurately (up to

2-4 weeks prior to the test), but does not indicate the level of parasitemia. The test can also

distinguish infections due to different Plasmodium species, but because almost all infections

in our sample were due to the most severe form of malaria (caused by P. falciparum), we only

present pooled results. Malaria prevalence is thus a cross-sectional estimate of the fraction of

tested individuals with the illness at a given point in time. Anemia, defined here as Hb levels

below 11 grams per deciliter of blood, is a common health condition in developing countries

and can be severely worsened by Malaria (White 2009). A significant change in anemia rates

in U5 is often one of the most sensitive indicators of changes in malaria incidence (Hawley

et al. 2003, ter Kuile et al. 2003). For each individual we also recorded respondent-diagnosed

illness episodes during the previous six months, which allows us to construct measures of

malaria incidence during the period. Unlike prevalence, this index is a ‘flow’ variable that

measures the overall burden of disease in the study population over a period of time. In

Section 4 we discuss at length the relative merit of these two indicators as well as their

relationship in epidemiological models of malaria transmission.

After the completion of the baseline, the 141 villages were randomly assigned to three

study groups of 47 villages each. We label the three arms (described in detail later) as “MF”

(microfinance), when nets were offered for sale on credit, “Free”, when the intervention called

5Appendix A.1 includes the details of the sample selection, and documents how study villages were on

average larger and with better amenities than the overall population of villages in the five districts.

5

for free distribution of ITNs, and “Control” when neither intervention was introduced. In

Table 1, we report selected summary statistics from the baseline, together with tests for

balance across treatment groups. The null of equality of means across arms is not rejected

at standard significance levels in 20 of 23 variables, suggesting overall good inter-arm balance,

although there are exceptions that we discuss below.

Average total expenditure was low, about 1.5 USD per person per day in purchasing power

parity terms, and approximately 20% of households were below the official poverty line for

rural Orissa (see the caption of Table 1 for details). Despite all sample households being

affiliated with BISWA, more than half said they would find it difficult or impossible to borrow

Rs 500, which is approximately the price of two program ITNs (see below). Two-thirds of

households had at least one net, 95% of which had been purchased from the market. The

mean (median) price paid was Rs 79 (60). The number of treated nets owned was significantly

lower, ranging from 0.02 ITNs per head in control areas to about 0.05 in Free and MF villages.

Despite the low ownership rates in all three arms, the null of equality is rejected at the 5%

level. More than 10% of tested individuals resulted positive for malaria, while more than

half were anemic.6 Malaria prevalence was marginally higher in treatment areas, 11.5% in

MF and 12.3% in Free, versus 10.8% in Control villages, although the differences are not

significant (p-value= .838). Self-reported malaria incidence was also higher in MF and Free

areas (0.12 episodes per person in the previous six months) relative to Control (0.09), and

in this case the null of equality is rejected at the 5% level. Estimated malaria-related health

expenditures were similarly higher in treatment areas, although the null of equality is not

rejected at standard levels (see table captions for details about the estimation of the malaria

costs). Overall, these rates document the poor health status of the study population and

suggests potentially large health gains from a reduction in the malaria burden.

In September-October 2007 we revisited the study villages and carried out a public infor-

mation campaign (IC), after gathering all BISWA members in a village. The IC included a

brief presentation about malaria and its transmission, the importance of ITN use, a demon-

stration of how to hang and use nets properly and advice on re-treatment. In treatment

communities, the IC also included an explanation of the intervention to be rolled out.

In the 47 Free villages, all households with at least one BISWA member (regardless of

inclusion in our baseline sample) received a number of free nets as a function of family com-

position, up to a maximum of four. The nets were of very good quality and significantly

6Malaria prevalence was similar across genders and age groups, while Hb levels vary widely by age and

gender (a common finding in developing countries), see Appendix A.3 for details.

6

sturdier than most of the pre-existing ones. They were treated with deltamethrin, an in-

secticide demonstrated to be effective against Anopheles mosquitoes in Orissa. Nets were

treated on the spot by trained personnel.7 Individuals were also informed that our team

would return after six and twelve months to re-treat the nets at no cost.

In MF communities, ITNs were offered through micro-loan contracts and, as in Free

communities, only BISWA clients were targeted. ITNs could also be purchased for cash.

The micro-consumer loans were offered by BISWA separately and in addition to any other

loan already outstanding. There was no movement of funds at the time of purchase: if a

household decided to buy ITNs, these were delivered after being treated as described above

and repayment was scheduled to be completed within one year. Field workers clarified that

default on ITN loans would be treated similarly to defaults for other BISWA loans and

that purchase decisions would not affect their access to regular BISWA loans beyond that

determined by repayment behavior.

ITN distribution and recording of loan contracts were to be completed 2-3 days after the

IC.8 The time interval between the IC and the purchase decision was introduced to ensure

that the households had an opportunity to consider the offer carefully. A second visit was

conducted approximately one month later, where ITNs were offered again with the same

contracts. No ITNs were offered after this second visit.

ITNs were offered for sale with two alternative loan contracts, both at BISWA’s standard

interest rate, 20% per year. The two contracts allowed buyers to choose between the purchase

of an ITN, or (for a higher price) a bundle which also included two re-treatments to be

completed at no additional cost six and 12 months later. The price of nets ranged from 173

to 259 Rupees, depending on contract choice or net size (single or double). For perspective,

at the time of the intervention daily wages for agricultural labor were around Rs 50, and

one kilogram of rice cost approximately Rs 10. Our project team re-visited MF and Free

villages in March-April and September-October 2008 for the re-treatment of the bednets,

which was completed by study personnel in a central location within villages. Re-treatment

was offered at no additional cost, except for buyers who did not choose the bundled contract

in MF areas who were offered re-treatment for cash, at Rs. 15 (18) per single (double) net.9

A detailed post-intervention survey was conducted shortly after the second re-treatment,

7See Appendix A.4 for specifics about bednets, insecticide, and the treatment procedure.8In reality, loan management was not carried out uniformly across the study areas by BISWA personnel.9Because of space constraints, in this paper we ignore the contract choice. Tarozzi et al. (2011) show

that, as expected, re-treatment rates were significantly lower among buyers who choose to purchase the ITNs

without the two re-treatments included in the price.

7

between December 2008 and April 2009. The content of the survey instrument was similar

to the baseline questionnaire and again measured ITN ownership and usage, and health

status. Malaria prevalence and Hb levels were measured by similar methodology to the

baseline survey. A longitudinal data set was created by re-contacting all baseline households

whenever possible. Additional funding also enabled us to increase the number of biomarkers

collected by attempting to test all household members for malaria and Hb, rather than only

specific demographic groups as at baseline.

Attrition at follow-up was limited and mostly due to temporary migration or inability to

find respondents despite repeated visits. Of the 1,844 initial households, 1,768 (96%) were re-

interviewed. The null of equal attrition rates among arms is not rejected at standard levels,

and neither bednet ownership nor the results of the biomarkers at baseline are statistically

or substantively significant predictors of attrition (see Appendix A.5 for details).

In describing the impacts of the interventions, we rely on intent-to-treat (ITT) estimates,

that is, we focus on post-intervention differences in outcomes between experimental arms

regardless of actual program uptake. We estimate all regressions using Ordinary Least

Squares (OLS), with statistical inference robust to intra-village correlation of residuals.

3 Impacts on ITN Ownership and Usage

We first evaluate the impact of the intervention on ITN uptake, ownership and usage. In

communities with free distribution, almost all sample households (96%) received at least one

ITN, with an average of 2.7 nets per household, about one every two people (Table 2, columns

1 and 2). In MF villages, ITN acquisition was substantively and statistically significantly

lower, with 309 of 589 sample households (52%) purchasing at least one ITN (1.2 nets

per household, or one ITN every four people). We also find considerable heterogeneity in

purchase rates across villages, with no uptake among sample households in five communities.

Almost all buyers chose to purchase on credit, with only ten choosing to pay in cash. Despite

the gap relative to free distribution, the 52% purchase rate was remarkable, given the non-

trivial cost of the ITNs. The high uptake contrasts sharply with the very low cash purchase

rates for health products documented among poor households in earlier studies such as Cohen

and Dupas (2010).10 In Section 3.2.1 we discuss the findings from an additional study where

we show that in a comparable set of BISWA communities demand for bednets offered only

for cash was very low and highly elastic with respect to price.

10Note also that the highest offer price for long-lasting ITNs in Cohen and Dupas (2010) was $1.35 (using

PPP conversion rates), that is, just above 10% of the least expensive ITN offered in our intervention. At

this low price, they estimate a purchase rate of approximately 40%.

8

Next, we assess the change in overall bednet ownership (regardless of acquisition mode

or treatment status of the nets) between the baseline and the follow-up survey. Column 3

of Table 2 shows the results of a differences-in-differences (DD) model where the dependent

variable is the change in the number of bednets owned by the household, and the regressors

are an intercept and dummies for households in MF and Free communities. We observe an

increase of 0.3 bednets per household in control areas but, consistent with the results on

ITN uptake, the overall increase was three times as large in MF communities, and six times

as large with free distribution.11 Free distribution led to a coverage of 0.63 nets per person,

which is close to the figure of two nets every three persons which has been taken to represent

full coverage in some contexts (see for instance ter Kuile et al. 2003).

The increase in net ownership in intervention areas was lower than the number of nets

delivered. The gap was on average 0.8 nets in Free and 0.3 nets in MF. In the latter

communities, the average gap is reduced to 0.1 if we exclude two outlier villages where a

number of BISWA members purchased more than 15 ITNs each for resale purposes. If we

exclude these two villages, in both Free and MF areas there was a 0.2 reduction in BISWA-

provided ITNs relative to the time of the intervention (results not shown). These ITNs had

been sold or otherwise lost or disposed of. In MF villages (again, excluding the two outliers),

we also observe a 0.1 increase in nets purchased from sources other than BISWA. Conversely,

in Free villages the additional 0.6 gap is explained by a decline in the number of non-BISWA

nets. This is consistent with the hypothesis that older, worn out nets had been disposed of

and replaced by the new high-quality ITNs distributed by our program.

Overall, we find that a large majority of ITNs distributed through the program were

retained. In addition, the surveyors were instructed to ask permission to see all nets that

the respondent listed as being owned by the household, and the presence of 90% of the nets

was confirmed in this way.

Information on bednet usage also confirmed large increases in intervention communities

relative to controls. Both at baseline and follow-up, we recorded whether household members

slept under a bednet the night before the interview, and whether the net had been treated

in the previous six months. In control areas, the proportion of members who slept under

a bednet changed from 13 to 18%, an increase likely due to the follow-up survey being

completed during a period of more intense mosquito activity (column 4 of Table 2). The

usage rate increased by an additional 9 percentage points (pp) in MF and 38 pp in Free

11Tarozzi et al. (2011) also includes, for all outcomes, the results of all regressions estimated in levels using

only information from the follow-up survey. Because observed characteristics were overall balanced across

arms, these estimates are always very similar and we do not report them for brevity.

9

communities. Two interesting patterns emerge when we look separately at changes of treated

and untreated nets (column 5 and 6). First, there was no increase in ITN usage rates in

control areas, which signals the absence of any cross-arm contamination due to imperfect

implementation of the study design or to the presence of other ITN distribution programs

in the area. Second, we find again evidence that the new, good quality ITNs supplied by

our program displaced non-treated nets, especially in areas with free distribution, where the

fraction of members who used an untreated net decreased by 8 pp relative to Control (the

decline is significant at the 1% level).12

Bednet usage during the previous night and the actual presence of the net in the dwelling

were also recorded independently in a census of sleeping spaces. Surveyors listed all sleeping

spaces used by the household (including outdoors), recorded which members slept there the

previous night, asked whether the space was protected by a net, and noted down the source

and price of the net and of any recent re-treatment. Surveyors asked to see all nets reported

as having been used. We use these alternative data to construct a new dummy for previous-

night usage of a treated net, and one for whether the net had been observed by the surveyor

and recognized as an ITN distributed through our program. The results are virtually identical

to the earlier ones, and while it is possible that misreporting was common to both sets of

responses, the remarkable degree of consistency across sections makes it unlikely.13 Even

though previous-night usage rates are likely a noisy indicator of consistent usage, the results

discussed so far show that the intervention increased ITN adoption substantially, but that

free distribution was much more successful than micro-loans at doing so.

A related question is whether the price of ITNs sold on credit generated a screening effect,

defined (as in Ashraf et al. 2010) as higher usage rates conditional on ownership, relative to

what observed with free distribution. This form of screening is often used as an argument

in favor of cost-sharing. Ashraf et al. (2010) find that households who agreed to purchase a

water purification product at higher prices were more likely to use the product, at least in the

short term, while Cohen and Dupas (2010) cannot reject the null that women who received

free ITNs were as likely to use them as others who paid subsidized but positive prices. In

12We also collected information on “regular” usage during the peak mosquito season. The seasonality

of malaria transmission has been documented in neighboring areas (Sahu et al. 2003, Sharma et al. 2006).

Regular usage rates are substantially higher than previous-night usage, but the cross-arm gradient is similar.

The results, omitted for brevity, are available in Tarozzi et al. (2011), where we also document that changes

in usage were very similar between genders, but larger for younger individuals, especially in Free areas.13In addition, such concordance is not simply due to all members being reported as either having or not

having used nets the night before. The correlation between the two separate reports is still very high (0.87)

if we use only information from households where there is intra-family variation in reported usage.

10

contrast, we find that while in MF areas 31% of ITNs had been used the night before, the

fraction was 14 pp higher in Free villages, and the difference is statistically significant at any

standard level, see column 8 of Table 2.

3.1 Correlates of ITN Purchases on Credit

In Table 3, we look at correlates of ITN purchases in MF villages. While these results are

descriptive and do not imply causal associations between the predictors and the decision to

purchase, they provide useful information on two key issues. The first is whether the sales

on credit led to selection into ownership of households with relatively high expected benefits

from ITNs. The second is whether purchase decisions are consistent with the presence of

credit and/or liquidity constraints, which would help rationalize the high uptake of ITNs sold

on credit. To analyze these points, we estimate a Linear Probability Model where the binary

dependent variable is equal to one if the household purchased at least one ITN (marginal

effects calculated from a probit model, not reported, are almost identical).

Variables that describe the demographic structure of the household (including presence

of U5s) are not significant, either individually or jointly (p-value= 0.6276). However, we find

strong associations between demand and proxy measures of perceived benefits from ITNs.

First, conditional on other covariates, households where everyone used a net prior to the

intervention were 21 pp more likely to purchase nets relative to others where no one did.

This is consistent with bednets being an experience good, with past usage perhaps associated

with higher perceived benefits (Dupas 2012b). Second, an increase from zero to the median

monetary cost of malaria episodes in the 6 months before the interview increased demand

by 9 pp (0.019 × 5901/4). Third, a history of any malaria-related deaths in the previous

five years increase the predicted probability of purchase by 10 pp. However, deaths were

rare (only nine respondents reported any) and the coefficient is not significant. Fourth, both

self-reported malaria episodes and prevalence as measured by our blood tests are among the

strongest predictors of purchase. Moving from a household with no self-reported malaria

incidence to one where every member had been sick increases the probability of purchase by

27 pp. Similarly, an increase from 0 to 100% in the fraction of blood tests administered to

the household that were positive for malaria predicts a 20 pp increase, and both coefficients

are significant at the 1% level. In contrast, we find that anemia levels are not correlated

with demand for ITNs, as in Cohen and Dupas (2010). This is consistent with anemia being

a poor indicator of perceived marginal benefit from ITNs, perhaps because among poor

households low Hb levels are often caused by a number of epidemiological and nutritional

factors besides exposure to malaria (de Benoist et al. 2008).

11

3.2 The Role of Liquidity Constraints in Demand for ITNs

The strong association between willingness to pay for ITNs and malaria risk overall indicates

that, despite the possibility of delayed payment, non-negligible ITN prices led to significant

selection. This finding suggests that credit and/or liquidity constraints (that is, lack of

ability to pay) were key factors in explaining the low ITN ownership rates observed at

baseline. Several pieces of evidence support this hypothesis. First, only 10 of the 309 buyers

in the sample chose the available option to purchase for cash. Second, households with lower

monthly expenditures were more likely to purchase ITNs, despite controlling for ownership

and usage of pre-existing nets: a 10% increase in per capita expenditure predicts a 1.2%

decrease in the probability of purchase, with the slope significant at the 5% level. Poorer

households may have found the opportunity to purchase ITNs on credit more appealing.

Third, we have shown that bednets were already present in the area, although few bednets

were treated and our ITNs were overall of better quality relative to those available in local

markets. Hence, high purchase rates were unlikely to be merely the result of ITNs being a

new product, not available outside of the intervention.

In principle, an alternative explanation for the purchases on credit was the presence

of alternative investment opportunities for their cash that yielded a return higher than the

BISWA interest rate (20% annually). However, in that case (and in the absence of investment

ceilings) one would have expected households to be maximizing their BISWA borrowing.

Although we cannot rule out this possibility completely, we find that only about 14% of

households had a current BISWA loan at follow up (excluding the ITN loan). Another

possibility is that the preference for purchase on credit relative to cash was due not to

liquidity constraints but to buyers having present-biased preferences. A purchase on credit

could have been seen as a way to start enjoying the benefits of ITNs while postponing the

associated costs. However, we find that an indicator of present-biased preferences predict

neither the decision to purchase nor the choice of cash vs. credit.14

To further probe the hypothesis that the relaxing of liquidity constraints were a crucial

factor leading to high demand, we conducted a follow up study between February and April

2012, where bednets were offered only for cash to BISWA households. The presence of

significantly lower levels of demand in this context would support our hypothesis. Although

the different timing means that we cannot rule out the possible role of time-specific factors

14The results are available upon request from the authors. The indicator is a measure of whether the

respondent exhibited “preference reversals” in a set of intertemporal choices, similarly to Ashraf et al.

(2006). See the caption of Table 3 for additional details.

12

on demand, we argue that a number of key factors were likely to a priori bias the results

against such a finding.

3.2.1 Design of a Follow-up Study of Cash Sales

We carried out the new intervention in 40 BISWA villages (“Cash” villages hereafter). Of

these, 25 were selected at random from the 47 previous control villages, where no ITN sales

had been conducted, while the additional 15 were newly sampled from our original sampling

frame. The 25 previous control villages (“PC”) were included because new villages (“New”)

were exposed to neither a comparably intense malaria and ITN-focused questionnaire nor

to blood tests. If these factors were important in increasing demand in MF villages, the

inclusion of only New villages might have biased results in favor of our hypothesis about the

centrality of liquidity constraints. Again in order to avoid biasing the results towards finding

low demand, the 40 sample villages were selected randomly after excluding communities

where BISWA was no longer operational and/or where public health programs or NGOs

had initiated bednet disbursal programs after the post-intervention survey in 2009. Due to

funding constraints, a household-level survey was not completed for this follow-up study.

However, data from the 2001 Census of India show that community-level characteristics in

the 40 Cash villages were overall very similar to those of the other villages originally included

in the study, and that they were also similar between PC and New villages, see Appendix

Table A.9 for details.

In cooperation with the micro-lender, field workers identified all members of BISWA self-

help groups in the 40 Cash villages. An information campaign similar to that in 2007 was

then conducted, discussing malaria and bednets, and describing the sale that would take

place in the following days. Each BISWA household was then provided a voucher, that is, a

slip of paper with the household’s name and the price of the ITN printed on it. The vouchers

were distributed to facilitate the calculation of the fraction of BISWA households purchasing

nets (the total number of vouchers distributed being the denominator). Then, as in MF

communities, sales were completed during two separate visits to the village scheduled in the

following days. During each visit, BISWA members who wanted to purchase bednets did so

by paying cash, on the spot, after returning the voucher.15

A key difficulty in generating comparability between the Cash and MF arm was that the

nets sold in 2007 required periodic re-treatment with insecticide to maintain efficacy. As

funding and timing issues did not allow us to schedule the re-visits 6 and 12 months after

15If the voucher had been lost, a new one was created on the spot and the purchase recorded.

13

the sale, we substituted the nets with OlysetTM long-lasting insecticidal bednets (LLINs).

In these nets, the insecticide permethrin is incorporated into the fabric itself, so periodic re-

treatment is not necessary. The use of these LLINs has been recommended for prevention of

malaria by the WHO since 2001 (WHO 2001). Olyset nets have been shown to maintain their

insecticidal properties even after four years in field conditions, thus guaranteeing significantly

longer protection relative to the program nets delivered at the time of initial distribution.

An additional advantage is that the mesh of these LLINs is wider than in most traditional

bednets, so their usage in hot weather causes less discomfort because of better air circulation.

Both the increased life-span and the wider mesh indicate that the LLINs were a higher quality

product than the ITNs sold in the 2007 intervention.

Although these LLINs were significantly more expensive than the ITNs marketed in 2007,

we priced them at a subsidized level to enhance comparability with the earlier prices in the

MF intervention. Potential buyers were informed about the market price of the LLINs (it

was also printed on the LLIN packaging) which was about twice as high as our offer prices.

The LLINs were then either sold at the same nominal price as the ITNs sold in 2007, or at

the same real price, calculated by inflating the nominal price in 2007 using a price index for

rural Orissa. Randomization of prices was done at the village level so that all households

in a given village were offered the LLINs at the same price. Census data show that village

characteristics were overall balanced in high and low price communities, see column 3 in

Appendix Table A.9.

On the one hand, the Cash arm involved the sale of better bednets at a real price either

below or identical to that relevant in 2007, with even the largest price being heavily subsidized

(and advertised as such). These factors would have likely increased demand relative to what

we would have observed had we implemented this Cash arm at the same time as the sales

on credit, biasing a comparison with sales on credit against our hypothesis. On the other

hand, the different timing of the cash sales relative to the original sale on credit implies that

we cannot control for confounding time-variant factors such as changes in malaria risk.

3.2.2 Demand for Bednets with Cash Sales

Vouchers were distributed to a total of 1,728 households, and 187 of these (10.8%) purchased

a total of 275 LLINs (0.159 per household on average, or 1.5 among buying households), see

Table 4 for details. As a reminder, when ITNs were offered on credit, we found that 52% of

BISWA households purchased at least one. The difference is significant at the 1% level.16

16Taking into account that the Cash and MF samples were independent, the t-ratio can be calculated

simply as (.52− .108)/√.052 + .0192 = 7.7.

14

Demand was very similar between PC and New villages, and we cannot reject the null

that the fraction of buyers was the same between the two (p-value= 0.72). We also find that

while 15% of households redeemed vouchers in villages where the LLINs were sold at the

lower price, demand was 50% lower when the price was increased by 20% relative to 2007

to take inflation into account, a difference that is statistically significant at the 10% level.

This implies a very high elasticity of demand equal to 2.5. This is important, because it

shows that households in our study areas were very sensitive to price variations when faced

with cash payments, consistent with the earlier literature that documented high elasticities

of demand for health products in samples of poor African households.17

One last interesting observation emerges by looking at demand in PC villages as a function

of whether the household had been included earlier in the sample (rows F and G). Among the

282 households who had been included in the baseline survey, 12% purchased LLINs. The

proportion was lower but similar (9%) among other households who had not been part of

the baseline survey. The null of equal demand between the two groups cannot be rejected at

standard levels. This is potentially important, because it suggests that the earlier exposure

to the survey and the RDTs had only a marginal impact on demand. In principle, exposure

to the IC, to the malaria-focused questionnaire and to the RDT results may have encouraged

ITN adoption regardless of the offer of delayed repayment. In Appendix A.6 we explore this

possibility more thoroughly, making use of data from two additional data sources: first,

the purchasing behavior in MF villages of BISWA households that had not been randomly

selected for inclusion in the baseline survey; second, information on ITN usage and attitudes

towards malaria from 25 additional villages surveyed at follow-up (in 2008-09), where no IC

or survey had been carried out earlier. Overall, we conclude that the IC was not a plausible

confounder, while exposure to the survey and RDTs may have increased demand in MF

villages but can only explain a small fraction of the relatively high uptake.

17The difference in demand between MF and Cash interventions is even more remarkable if one takes into

account the way we measured take-up. In the MF arm, this was measured as a ratio where the denominator

was the number of households included in the baseline survey, conducted a few months before the IC and

the sale, while the numerator was the number who purchased at least one ITN. In contrast, in the Cash

arm the denominator was the number of households who received a voucher during the IC. So, a BISWA

household who did not attend the IC/sale and hence did not purchase any net would have been included in the

denominator and counted as not buying in MF villages, but it would have been excluded from the calculation

in Cash villages, thereby biasing estimated demand for cash upwards relative to MF. If we estimate demand

by replacing the number of households who received vouchers in the denominator (1,728) with the number

of BISWA households in the 40 villages listed in the rosters provided by the micro-lender (1,971), overall

demand declines from 10.8% to 9.5%.

15

4 Malaria Indicators: Descriptions and Impacts

Next, we analyze whether the large increase in ITN ownership and usage in treatment areas

was reflected in improvements in malaria indices, namely malaria prevalence and incidence

and Hb levels. Before doing so, we clarify the nature, inter-relationship and measurement of

the two most direct indicators of malaria burden: prevalence and incidence. This is impor-

tant, both because the two indicators are not equally good measures of the direct economic

burden of malaria, and because empirical findings and epidemiological models of malaria

transmission suggest that they may not respond equally to public health interventions.

First, recall that in our data malaria prevalence measures the fraction of tested individ-

uals with the disease at a given point in time, estimated from rapid diagnostic blood tests

(RDTs). This is a key malaria index, also because it represents the frequency of individuals

who may transmit the disease to others through Anopheles bites.18 However, in areas of

intense transmission such as our study locations, individuals who test positive are frequently

asymptomatic due to partial immunity acquired from repeated infections (Laishram et al.

2012).19 Accepted epidemiological models are consistent with acute episodes of malaria oc-

curring only when a host experiences either high parasite density or in the presence of other

risk factors (e.g. Ross et al. 2006). Although asymptomatic cases of malaria may not lead to

significant direct costs to households, they remain of great concern for public health, because

they complicate considerably attempts to eradicate or mitigate the disease.

Despite this, malaria incidence, that is, the total number of infections per person in the

study population over a period of time, is often considered to be a more comprehensive mea-

sure of disease burden than prevalence. Measuring incidence accurately, however, requires

repeated, regular re-assessments over short periods of time and is therefore expensive and

invasive. Indeed, six of the 22 ITN trials reviewed in Lengeler (2004) measured only preva-

lence. In our study malaria incidence was estimated from detailed recall information about

illness of household members during the previous six months.20 We recorded all malaria as

well as fever episodes that led to absence from work or school, or to consultation with health

workers or hospitalization, noting all the related monetary costs as well as the number of

days of work or school lost, see the caption of Table 6 for details.

18In addition, the completion of the life cycle of Plasmodium requires the infection of a host, so that

malaria cannot spread in mosquitoes alone (White 2009).19For instance, McMorrow et al. (2011, Fig. 1), using data from Malaria Indicator Surveys in 2007-2009,

show that among children in Kenya, Mozambique, Senegal and Zambia who tested positive the ratio of

asymptomatic to symptomatic ranged between 1.4 and 6.9.20Pilots suggested that longer recalls led to significant respondent fatigue.

16

Self-reports are likely affected by recall error and misdiagnosis that may be non-random,

see Strauss and Thomas (1998, Section 4). Das et al. (2012) found that longer recall periods

led to lower reported morbidity per unit of time in a sample of individuals in Delhi, India,

and that the ability to recall was correlated with socio-economic status. In our sample, less

than 1% of individuals were reported as having had malaria within a month of the post-

intervention survey, while RDTs detected the presence of the malaria parasite in 21% of the

tested individuals (see below). Errors of inclusion were also common, given that only 28 of

the 63 individuals reported to have malaria tested positive with the RDT. In principle, errors

of recall or diagnosis may have been correlated with treatment, with unclear implications

for bias: for instance, the distribution of ITNs may have made the disease more salient,

pushing respondents to over-report illnesses, or it may have led to a decrease in the perceived

malaria risk, with opposite effects. Despite these limitations, in Appendix A.7 we show that

respondent-diagnosed recent malaria episodes were strongly correlated with the results of

the RDTs, and that this association was not differential across experimental arms. Overall,

this suggests that self-reported incidence was thus a valuable if imperfect indicator of past

malaria episodes. In addition, one advantage of respondent-diagnosed illness episodes is that

they likely identify (unlike RDT-based prevalence) illnesses that were severe enough to be

recognized by the household, and thus potentially more important from the viewpoint of the

economic burden they imposed on the household.

In epidemiological models, prevalence and incidence are strictly linked by a relationship

that depends on frequency of infection and recovery time. However, the two indicators

are distinct and they may respond differently to health interventions, even when both are

accurately measured with blood tests. For instance, Beier et al. (1999) show that significant

reductions in prevalence are usually achieved only with large reductions in the entomological

inoculation rate (EIR, the number of infective bites per person/year), an indicator strictly

related to incidence.21 Intuitively, an anti-malaria intervention may succeed in reducing

the number of infective bites (the intensive margin) while barely affecting the probability

of receiving some infective bites (the extensive margin). Indeed, substantive differences in

estimated impacts of ITNs on malaria prevalence vs. incidence have been found in several

earlier studies. Among the 22 ITN impact studies reviewed in detail in Lengeler (2004), only

seven measured both prevalence and incidence of malaria, but in all those cases the protective

power of ITNs was found to be larger when looking at the latter. Two studies actually found

21As few as 1-10 infective bites per person/year have been associated with prevalence rates ranging from

about 10 to 80%, Beier et al. (1999, Fig. 2). In malaria-endemic areas, EIR above 100 are common. In

locations close to our study districts, Sharma et al. (2006) documented EIR in the range of 3-114.

17

substantial improvements in P falciparum incidence while documenting higher prevalence in

treated areas, although the increases were not significant at standard levels.

Empirical studies and epidemiological models of malaria transmission and ITN usage

also suggest that the malaria burden is best reduced when a large fraction of the population

has access to ITNs and when the nets are used regularly. Regular usage provides private

benefits by limiting the number of infective bites, but a high ITN coverage rate can also

be key, when it leads to substantial externalities achieved through declines in the number

of mosquitoes (Binka et al. 1998, Hawley et al. 2003). In particular, Killeen et al. (2007)

describe a rich epidemiological model calibrated using data from a number of field studies.

One of their key results is that a user protected for 90% of the time is predicted to reduce the

EIR by 60% relative to a non-user if no other ITNs are used around him, but the reduction

becomes close to 100% if everyone else in the community is also regularly using ITNs. In the

previous section we have demonstrated that ITN usage increased substantially in our study

areas, especially with free distribution, although usage rates remained low, with less than

half of ITNs reported in use the previous night. In addition, our ITN distribution programs

only targeted BISWA households, so that even in Free villages ITN coverage remained rela-

tively low (about 20% on average). These factors may thus have limited the impact of our

interventions, especially on malaria prevalence, and especially in MF areas, where less-than-

universal uptake limited ITN coverage and where we also recorded, surprisingly, usage rates

lower than in Free areas, even conditional on ownership.

4.1 Results

We first look at the data based on RDTs, that is, malaria prevalence and Hb levels. As a

reminder, at follow-up all members of sample households were targeted for testing of both

malaria and Hb, while only a subset were at baseline. At follow-up, 75% of individuals were

successfully tested, while 19% were not because they were absent during the visits and 6%

refused. Both refusal and absence were balanced across experimental arms (see Appendix

A.8). The ITT estimates of the program impact on RDT results are reported in Table 5,

where we show results of regressions both in levels (using all tests completed at follow-up)

and in DD form (for the fewer individuals tested in both surveys).

At follow-up malaria prevalence was 18.3% in control areas, 22.7% in Free and 22% in MF

communities (column 1). Malaria prevalence was therefore about 20% higher in intervention

communities, although the null of no difference between each intervention arm and control

areas cannot be rejected at standard levels.22 The estimates are sufficiently precise that we

22Given that ITN ownership and usage are higher in Free and MF villages relative to controls, these results

18

can also reject the null hypothesis of large reductions in malaria prevalence in intervention

relative to control areas. The lower bound of the 95% confidence interval for the difference

between Free and control is −0.022, which corresponds to a 12% reduction in prevalence

relative to control areas. Similarly, the corresponding lower bound for the difference between

MF and controls (−0.025) would imply a 13% lower prevalence than in control villages.

Several earlier RCTs evaluating the impacts of ITN adoption found reductions in prevalence

substantially larger than these lower bounds (see Lengeler 2004, Appendix 8 and 9), although

we have discussed above that some studies found no improvements.

The higher prevalence in Free and MF areas could have been explained in part by pre-

intervention differences. The figures in Table 1 show that before the intervention malaria

prevalence in Free and MF villages was respectively 7% and 14% higher relative to control

areas, although the differences were not significant at standard levels. However, the DD

estimates, which only include individuals tested both before and after the intervention, are

similar to the results in levels (Table 5, column 2). Relative to baseline, malaria prevalence

in control areas increased from 11 to 17.3%. The overall increase in prevalence was expected,

because the baseline survey was completed during the hot and dry months of spring, when

malaria prevalence is lower, and the follow-up survey during winter, when malaria prevalence

is generally higher in Orissa (Sharma et al. 2006). Consistent with the results in levels, the

increase in prevalence was 5 pp higher in Free communities and 6 pp higher in the MF arm,

although again the differences are not significant at standard levels.23

Looking now at Hb levels, when we use all follow-up data, mean Hb levels was 11.4 g/dl in

control and Free villages, and 11.5 in MF communities. The estimated impacts are therefore

close to zero and not significant at standard levels (column 3). When we look at the DD

estimates, we find that mean Hb increased by 0.28 g/dl in control areas, 0.32 in MF and

0.50 in Free villages.24 The DD between Free and control areas is then 0.22 g/dl, or about

14% of a (baseline) standard deviation and is significant at the 5% level. The magnitude

is small but not negligible. For perspective, among the nine ITN efficacy trials reviewed in

Lengeler (2004) that measured Hb, impacts ranged from 0.2 to 1.8 g/dl, with a mean impact

of 0.67 g/dl although D’Alessandro et al. 1995, an effectiveness study such as ours, found

also lead to a positive association between malaria prevalence and ITN usage or ownership, if one estimates

the relationship with instrumental variables using treatment status as instrument.23When we calculate mean changes in malaria prevalence within villages, we find that prevalence declined

in only 11 of 47 control, 9 of 47 Free and 8 of 47 MF villages, while we observe increases in prevalence in 20

control, 27 Free and 30 MF communities, and no change in the remaining locations.24The increases in Hb, despite the higher malaria rate, was perhaps due to better nutrition at follow-up,

conducted in months when our data indicate that income was seasonally higher for many households.

19

an impact substantively lower than in our case (0.1 g/dl).25 Thomas et al. (2006) find that

an iron-supplementation program that specifically aimed at reducing anemia rates increased

Hb levels by 0.18 g/dl among adult males, and 0.12 g/dl among adult females.

When we look at anemia prevalence, defined as Hb< 11g/dl, we find that it was 38.4%

in control areas, 39.4% in Free and 38.9% in MF villages (column 5). Anemia was thus close

to identical across arms, with differences not significant at standard levels. The DD results

show similar patterns (column 6). The relative improvement in mean Hb in Free villages is

reflected in a 2.4 pp reduction in anemia relative to control, but in this case the DD is not

significant at standard levels. We also find that the lack of improvements in malaria and

anemia prevalence was common to all demographic groups (see Appendix A.9 for details).

In principle, the absence of any improvement in malaria prevalence may also have been

caused by measurement error, but this is unlikely in our context. First, random misclassifi-

cation of a binary dependent variable leads, by construction, to negative correlation between

the error and the true value of the variable. As long as the true and the mis-measured

values are positively correlated (as they likely are in our case) this leads to attenuation bias

(Hausman et al. 1998, eq. 15). As prevalence tended to be higher in treatment areas, mis-

classification would more likely have led to underestimation of the differences. Second, at the

beginning of the study, the reliability of the RDTs was successfully checked by testing a lim-

ited number of blood samples with or without malaria infection. On the other hand, during

the field work RDT results were not confirmed with microscopy and a degree of subjectivity

does exist in interpreting the results of the RDTs, which are read on a test strip located

on a card where a reagent is added to the blood sample. The presence of recent infection

with Plasmodium is signaled by the appearance of darker lines on the white strip. Although

high concurrency between test readers (including non-trained ones) has been documented in

clinical trials of the RDT (see Appendix A.2), a degree of subjectivity is hard to rule out

completely, because the lines can sometimes be difficult to detect when parasitemia is low.

In addition, if parasitemia was declining in treatment villages over the course of the study,

the likelihood of fainter, harder-to-detect test lines may have increased in these areas, which

would most plausibly have led to overestimating the reduction in prevalence.

To probe further the degree of subjectivity in our context, we carried out a small val-

idation study in collaboration with the Malaria Research Centre (MRC) Field Station in

Rourkela (Orissa). The results showed very high sensitivity (> 90%) and specificity (74 to

25Several studies report the results as ‘packed cell volume’. This can be estimated by multiplying by three

the Hb level expressed in grams per deciliter of blood.

20

85%), see Appendix A.10 for details. Further, we checked whether systematic differences in

the interpretation of the malaria RDT played a role in the results by re-estimating program

impacts with the inclusion of tester fixed effects (see columns 7 and 8 of Table 5). The

differences among experimental arms become slightly smaller, but they remain positive and

not significant at standard levels.

Finally, measurement error was unlikely to be a problem for the Hb testing, which also

showed mixed evidence of differential changes across experimental groups. Although erro-

neous testing cannot be ruled out entirely, measuring Hb simply requires reading a number

from the display of a small diagnostic piece of equipment. In addition, the strong cross-

sectional correlation between malaria infection status and Hb levels supports the reliability

of the malaria RDTs. When we regress Hb on a dummy for a positive malaria test, the slope

(= −0.19) is significant at the 1% level.

We also evaluate the hypothesis that malaria indices remained high due to behavioral

changes that may have compensated for the benefits of ITNs. The increased availability of

ITNs in Free and MF villages may have reduced the use of alternative prophylactic measures

such as indoor or outdoor wall spraying with insecticide, mosquito coils, or the control of

drainage pools. We tested this hypothesis using data on knowledge and practices collected

during the post-intervention survey, but the differences between arms are generally small

and show no systematic pattern, see Appendix A.11 for details. A potentially important

exception is wall spraying which, like ITNs, is widely considered an effective mean of reducing

malaria risk (World Health Organization 2007). In control areas, 40% of households had

the inner walls sprayed after 2007, while the proportion was 37% in Free and 30% in MF

communities. The proportions who had the outer walls sprayed in the three groups were

respectively 53, 48 and 44%. Although the null of equality is not rejected, the magnitude

of the differences is relatively large. We then re-estimate the ITT including dummies for

recent wall spraying among the regressors. Spraying is potentially endogenous, but here we

are only interested in evaluating whether differences in spraying rates help explain the lack

of health benefits in intervention villages. In columns 9 and 10 of Table 5 we show that

this leaves the estimated impacts on malaria prevalence almost identical (the results for Hb

are similar too and are available upon request). Overall, then, we find no evidence that our

results are due to changes in household risk-coping behavior.

Similarly, the lack of effect on malaria or anemia prevalence cannot be explained by the

presence of other ITN distribution programs, possibly sponsored by the Government or by

other NGOs. First, the results on net ownership in Table 2, which showed large increases

21

in ITN ownership rates in treatment versus control areas, included nets from all sources.

Second, we find that the number of nets received from non-BISWA sources was very small

and not significantly different across all arms, see Table A.12 in the appendix for details.

An additional concern is the possibility that malaria and anemia prevalence did not im-

prove either because the bednets had not been treated appropriately with deltamethrin (the

insecticide), or because the local population of Anopheles mosquitoes was or became resis-

tant to the chemical. We cannot address these concerns directly, because our data include

neither systematic measurements of the insecticide concentration on the ITNs nor informa-

tion on number, behavior and susceptibility to insecticide of local Anopheles. We argue,

however, that these factors were unlikely to be central. First, all bednet treatments were

conducted by our trained personnel, using appropriate procedures and chemical concentra-

tions, and tests run on a small number of ITNs at the end of the study were consistent

with adequate treatment, see Appendix A.4 for details. Second, none of a number of recent

studies carried out in Orissa and Madhya Pradesh (another Indian state) point to ITNs as

a plausible cause of resistance to deltamethrin, see Appendix A.12. In addition, resistance

to insecticides is unlikely to develop over a relatively short period of time in a situation such

as ours, where ITN use was largely limited to a study population (BISWA members) that

always represented a minority of the village.

Next, we analyze our data on respondent-reported malaria incidence. Although recall

error make these data less reliable than RDTs, we have argued above that (abstracting from

such concerns) incidence is perhaps the best measure of disease burden. In Table 6 we show

the estimated ITT program impacts on a number of self-reported malaria indicators. We only

discuss the DD estimates because, unlike the RDTs, some of the self-reported outcomes were

not balanced at baseline, suggesting higher malaria burden in Free and MF areas relative to

Control, see Table 1. We first look at the fraction of individuals with episodes of malaria that

are still ongoing or recent (within a month). These figures can be interpreted as self-reported

prevalence and (as we noted before) are remarkably low, likely suggesting that most malaria

cases identified by RDTs were asymptomatic.

In Control areas only 0.7% of individuals were reported as having had malaria during the

last month and, consistent with the RDT-based prevalence results, the null of equality across

arms cannot be rejected. In contrast, measures of incidence over the previous six months

show beneficial impacts of ITNs, both in Free and in MF villages. In control areas, mean

incidence at follow-up was 0.115 episodes per person over six months, an increase of 0.025

relative to baseline likely due to the seasonal pattern of the disease. However, the DD for

22

both Free and MF indicate a relative decline in incidence of about the same magnitude, and

the coefficient is significant at the 5% level in Free villages, and at the 1% level in MF. The

magnitude of these impacts is large relative to control conditions, with a relative risk ratio

of 0.56 = (.115− .051)/.115, broadly consistent with the reduction in uncomplicated clinical

episodes observed in earlier efficacy studies, see Lengeler (2004, Appendix 6,7). In addition,

and given that the average household had about 5.5 members, these estimates indicate a

relative decline of 2× 0.05× 5.5 = 0.55 episodes of malaria per household per year.

An important caveat is that a fraction of these malaria cases were in fact likely misdiag-

nosed fever episodes. Recall that we find that RDT results and respondent-diagnosed recent

malaria cases only coincided in 44% of cases, with similar rates of concurrence among exper-

imental arms (see also Appendix A.7). The actual impact on symptomatic malaria incidence

was thus likely smaller than suggested by the results in column 2 of Table 6. In Appendix

A.13 we show that the estimates are about 40% smaller (but still significant at standard

levels) if we assume that only 44% of incidence was correctly diagnosed, while also consid-

ering that a fraction of respondent-reported fever cases were likely misdiagnosed cases of

symptomatic malaria. Regardless of this adjustment, the estimates are substantively large,

given that at baseline the average monetary cost of a malaria episode was close to Rs 1000,

about 30% of total monthly household expenditure. The estimates are also large relative to

the price of our program ITNs, which cost at most Rs 259. Note also that the estimates are

ITT, surely a lower bound of the average benefit for the treated.

The changes in incidence were also reflected in the costs borne by households due to

self-reported malaria cases (columns 3-7). In both arms, the DD show an average reduction

per household of about two days of school or work lost due to malaria, relative to an endline

average of 5.8 days in Control areas. In control areas, we estimate that malaria cost Rs

863 per household during the previous six months, about 3% of total household expenditure

during the same period. Relative to Control areas, free distribution of ITNs was associated

with a Rs 194 lower expenditure due to malaria, while the reduction was Rs 269 in MF

villages, although only the latter is significant, at the 10% level. Such reductions correspond

respectively to 23% and 31% of estimates of total malaria costs over a six month period in

control areas at endline.26 The figures in column 5 show that a large parts of these costs

26The estimates in columns 3-7 (as those in column 2) do not distinguish between correctly self-diagnosed

malaria cases and other sickness episodes incorrectly diagnosed as malaria. We are interested in impacts

on costs regardless of whether they were actually due to malaria because respondents sought treatment

based on their perceptions (hence perceptions are what matters) and the ITNs may plausibly have led to

improvements in overall health as well due to decreased malaria burden. Note also that the impacts on

23

were due to doctors and drugs. Finally, we find that the interventions were associated with a

relative decline of about 0.1 in the number of malaria episodes that forced the households to

incur debt or reduce consumption to cope with the necessary costs. These DD are significant

at the 10% or below and are large relative to the levels in Control villages.

4.2 Interpretation and Discussion

Overall, self-reported malaria incidence rates are thus in stark contrast with prevalence as

measured with RDTs, both in terms of levels and in terms of changes over time. Reporting

error on malaria incidence correlated with the intervention could explain at least in part

these findings, but as we described earlier substantive differences in impacts on prevalence vs.

incidence are not unique to our study. Our findings are thus consistent with the hypothesis

that the increase in ITN usage reduced infective bites enough to reduce the case-incidence

of acute malarial episodes, but not enough to reduce the overall prevalence of malaria.

Low usage and coverage rates may have contributed to these findings. First, despite the

substantive increases in ITN usage documented in Table 2, only 45% of the program ITNs

were in use the night before the follow-up survey in Free areas, and about one third fewer

were in use in MF villages. Second, it must be recalled that only BISWA clients received free

ITNs or the offer of ITNs for sale on credit. Although BISWA had a large presence in the

study area, we estimate that on average only 20% of people lived in households with at least

one BISWA affiliate and thus were eligible for inclusion in the study. Even in villages where

nets were distributed for free, ITN coverage was therefore low, nowhere larger than 50% and

with only four villages where it surpassed 30%. With a 20% coverage and a frequency of

usage equal to the 45% cross-sectional usage rate, the epidemiological model of Killeen et al.

(2007) (leaving all other calibrated parameters unchanged) predicts a reduction in infective

bites of about 40% for users and significantly less for non-users, see Appendix A.14 for details.

Such declines, while substantive, may have been too small to be detected by measures of

prevalence (Beier et al. 1999). Of course these calibrations have to be taken with caution,

because the model in Killeen et al. (2007) depends on a number of parameters—such as

species, number and feeding habit of Anopheles—that we do not observe in our data.

Our data are not ideal to study the role of community-level ITN coverage on malaria

prevalence, both because the fraction of the population treated in intervention villages

(BISWA households) was always small and because random assignment of treatment was

malaria and fever pooled together are qualitatively similar to those for malaria only but the magnitudes are

larger (see Appendix Table A.14 for details).

24

not stratified by BISWA degree of presence in the village. Perhaps for these reasons, we find

no clear association between changes in prevalence and estimates of ITN coverage achieved

via free distribution, see Appendix A.15 for details. We also investigate the extent of within-

village externalities, by using geo-coded data from a sub-sample of 11 villages, four Control

and seven Free. We estimate a model where malaria status is regressed on the number of

total and BISWA neighbors within different radii, both interacted with the Free dummy.

The point estimates suggest substantial externalities at short ranges, with lower prevalence

in Free villages associated with increases in the number of BISWA neighbors within 5-20

meters (a proxy for local ITN coverage). This is consistent with the existence of clusters of

ITN-related protection, and with the argument that higher coverage rates may have led to

declines in prevalence. However, the small sample leads to very imprecise estimates that are

not significant, see Appendix A.16 for details.

To help rationalize the lack of improvement in malaria prevalence, it is also useful to

compare our study design with that of the 22 ITN efficacy studies reviewed in Lengeler

(2004). Fourteen were, like ours, clustered randomized trials, while in the remaining eight

ITNs were randomly assigned within community. Among the clustered RCTs, the largest

impacts of ITNs were found where community-level coverage was very high. In all trials, the

number of ITNs distributed was sufficient to ensure that a majority of sleeping spaces were

protected by nets in treatment communities. In addition, six of the seven clustered RCTs

that measured impacts on malaria prevalence achieved close to universal coverage. The one

exception is D’Alessandro et al. (1995), which is also the only ‘effectiveness’ study surveyed

in Lengeler (2004). That is, while all other studies evaluated benefits of ITNs under ideal

trial conditions (‘efficacy’), this study focused on sentinel sites for the evaluation of a public

health program in The Gambia. After one year, they observe substantial improvements in

malaria indices among children in treated areas. However, they also show that prevalence

was actually higher (71 vs. 45%) in treatment relative to control areas in one of the five

sentinel sites, despite no evidence of resistance to insecticide. As a likely key explanation

they mention “low usage of nets by children in this area” (p. 482). More evidence about

the importance of high coverage is also found in Kroeger et al. (1999), a study carried out in

Nicaragua where the fraction of individuals reported as sleeping under ITNs ranged between

5 and 70%. They find that declines in incidence were smaller in areas with lower coverage,

with no improvements detected when coverage was < 16%.

The studies with intra-community assignment of ITNs (and hence low community-wide

coverage rates) found 40-60% declines in malaria incidence. However, as in our case, preva-

25

lence increased in treated areas in two of the studies. Moreover, in these studies usage rates

were reported to be very high (70% or above), while in our context less than half of the

ITNs were reported as being used the previous night.27 It is possible that such high usage

rates were achieved because these studies (unlike ours) involved intense monitoring of net

usage and/or health outcomes, including a combinations of nightly surprise visits and fre-

quent (sometimes daily) health checks. Such a study design could have induced behavioral

responses such as increased compliance with regular ITN usage.

In sum, taking both the biomarkers and the self-reports as broadly correct, we conclude

that our intervention reduced the incidence of severe malaria cases sufficiently to lead to

declines in malaria-related expenditures that were large relative to control conditions. On

the other hand, we also found no improvements in prevalence, and comparisons with the

existing literature on ITN efficacy and malaria epidemiology suggest that this was likely the

result of relatively low usage rates and population coverage.

5 Cost Effectiveness Analysis

The cost-effectiveness comparison between free distribution and sales on credit would be

trivially in favor of micro-loans if repayments were complete and take-up comparable between

arms, but neither condition held in reality. Together with the cost of the ITN themselves

(which included delivery at the BISWA headquarters in Sambalpur), distribution expenses

included costs for labor and for transporting the ITNs from Sambalpur to the villages. Based

on field expenses, we estimate a transport cost of Rs. 500 per day and wages of Rs. 150 per

day per worker, and about 1.5 days to cover a village. If our intervention were scaled up

through a micro-finance network, these delivery costs could be lower if the delivery operations

were scheduled using the MFI’s existing labor and transportation resources. Insecticide

treatment costs were Rs. 10-13 per net (depending upon size). Dividing the total cost thus

obtained by the number of ITNs distributed in each arm lead to a cost per ITN delivered of

Rs. 305 in MF villages and Rs. 225 with free provision.

Turning to revenues, recall that the price charged to buyers already covered BISWA’s

costs in administering the loan. However, at the time of the follow-up survey, about 1.5

years after the sale, sample households in MF communities had repaid on average 64% of

ITNs, and we assume that no further payments were made afterwards. The low repayment

rates were largely due to some BISWA program officers not putting effort into enforcing

27Unlike our project, all these eight studies were also carried out within relatively small geographical areas,

with the exception of one where the study population was spread across one district.

26

repayments, especially in certain districts. This was despite what was conveyed to BISWA

members at the time of the sales, when they were informed that loan defaults would have

been treated as for any other BISWA loan. Note that our data are not consistent with

households anticipating that repayment would be scarcely enforced in some areas. This

is important, because if households had anticipated such enforcement behavior, the high

demand for ITNs on loan could have been a mere by-product of ‘sales’ actually perceived

as free or highly subsidized distribution. However, we find a positive correlation between

demand for ITNs and share of the loan repaid across study districts.

Overall, there are thus two key drivers of the differences in cost per ITN in the two arms.

On the one hand, about four times as many ITNs were delivered in Free villages, thus lowering

considerably the incidence of fixed costs per ITN distributed. On the other hand, in MF

villages a substantial fraction of costs were recouped through repayments. These calculations

yield a cost per ITN of about Rs. 150 in MF villages and Rs. 225 in Free villages. Sales on

credit, despite repayment and fixed cost concerns, were thus considerably more cost-effective

in the sense that for a fixed budget, 50% (=100 × [(1/150) − (1/225)]/(1/225)) more nets

can be distributed relative to free disbursal. However, given that purchase rates were well

below 100%, covering the same number of households under MF would require ITNs to be

distributed across more villages. For instance, the 4,000 nets that we distributed in total in

the 47 Free communities would have required reaching about 200 villages if they had to be

sold on credit. An important corollary is that cost effectiveness was achieved at the expense

of significantly lower within-village ownership rates relative to free distribution. To the

extent that externalities from mass distribution are an important source of ITN protective

efficacy, cost-recovery may be suboptimal since it will likely result in lower ITN densities.

An alternative way to evaluate the two delivery schemes is to consider their relative cost

of reaching “high benefit” households, defined crudely as those where any of the members

had malaria at baseline. By this metric, 60% of all households were high benefit. In MF

villages, such households were 17 pp more likely to purchase an ITN. These considerations

would tilt cost-effectiveness further in favor of microloans, implying that the cost of reaching

a high benefit household using microloans was Rs 227 and the corresponding figure with free

distribution was Rs 375, that is, two-thirds higher.

6 Discussion, Limitations and Conclusions

Liquidity constraints have been hypothesized to be a key reason behind the low adoption

rates of beneficial preventative health products among the poor in developing countries.

27

In this paper, we implement a randomized controlled trial to argue that micro-consumer

loans may provide a feasible and cost-effective method to increase adoption in situations

where existing markets and public health interventions have not been successful at ensuring

adequate coverage of ITNs, which are one of the most efficacious malaria prevention methods.

In a treatment arm composed of 47 villages in rural Orissa (India), our program succeeded

in selling about 1,100 ITNs on credit to clients of a micro-lender over a few months, despite

the relatively high price of the ITNs, about 3-5 times the daily agricultural wage in the

study area. This increased ITN ownership substantially relative to control areas, with 52%

of sample households purchasing at least one net.

These purchase rates are substantially higher than in earlier studies that found very

low cash purchases of health products among the poor, despite heavy subsidization (Ashraf

et al. 2010, Cohen and Dupas 2010, Kremer and Miguel 2007 and Kremer et al. 2009).

Consistent with these studies, we also find that the demand for bednets in our study areas

was significantly lower (11%) and highly price-elastic when households had to pay upfront

in cash. However, cash sales were conducted at a later time with respect to the sales on

credit, so we cannot exclude the presence of time-varying confounders. On the other hand,

we have described a number of factors that were likely to bias demand for ITNs offered for

cash upwards relative to what we observed in the sales on credit. We also provide additional

evidence that liquidity constraints played a key role in explaining the high adoption rates

of ITNs sold on credit, including the fact that only a handful of buyers chose to pay cash,

despite the option being available at the time of the sales on credit.

We also found clear evidence of selection into purchase, with indicators of past exposure

to malaria strongly associated with demand. In contrast, in a sample of women in rural

Kenya, Cohen and Dupas (2010) found no correlation between low levels of Hb levels and

willingness to pay for ITNs. This is possibly due to Hb being a noisy indicator of malaria

exposure, and indeed we also find no association. However, an alternative hypothesis is that

demand among at-risk women was reduced by positive correlation between malaria risk and

liquidity constraints, that is, by a negative correlation between willingness to pay and ability

to do so. In our setting, liquidity constraints were relaxed by the loan offer, and so correlates

of demand were less likely to confound willingness and ability to pay.

Despite a two-thirds repayment rate of loan at the time of the follow-up survey, 1 to 1.5

years after the sales, we estimate that sales on credit reduced the estimated cost of reaching a

household at risk (defined as one where any of the members had malaria at baseline) by about

50% relative to free distribution. Such considerations may be important for public health

28

programs that aim at maximizing the number of at-risk beneficiaries within the constraints of

a given budget. In situations where funding is only sufficient to offer protection to high-risk

individuals—such as pregnant women and young children—micro-loans may perhaps help in

approaching universal coverage by increasing adoption among individuals—such as working

adults—for whom episodes of clinical malaria may still lead to substantive economic costs.

However, these factors must be weighted against the lower product coverage achievable with

cost-sharing relative to free distribution, even when the product is offered on credit. This is

a potentially serious drawback in the presence of externalities.

When we estimate the program impacts on malaria indices, the results are mixed. First,

we find no evidence of substantial improvements in malaria or anemia prevalence (the fraction

of the population affected by the condition) when measured from blood samples. We find

an improvement (significant at the 5% level) in mean Hb in areas with free distribution, but

only when we use differences-in-differences estimates. In contrast, we find substantial and

statistically significant improvements in malaria incidence (the number of cases over a period

of time) in areas were nets were either donated or sold on credit, although these results are

based on recall data and were not clinically validated. Incidence was thus surely measured

with error, and our data suggest that a large majority of malaria cases detected with RDTs

were asymptomatic, a common finding in malaria endemic areas (such as Orissa) where

repeated exposure to the disease generates partial immunity. Fever cases misdiagnosed as

malaria were also likely common, although we have argued that recall data remain a useful

if biased proxy of symptomatic malaria cases, severe enough to be recognized within the

household.

The relative reduction in malaria incidence was also associated with lower health ex-

penditures, fewer days of work or school lost due to malaria, and fewer episodes forcing the

household to incur debt or lower consumption to pay for the related costs. The ITT estimates

suggest an average yearly saving in malaria-related health expenditure about twice as large

as the most expensive ITNs sold through our program. Given that the average impacts on

the treated was likely higher, and that our high quality ITNs should have lasted at least 2-3

years, these figures suggest that liquidity constraints imposed substantial health-associated

costs on the households. In addition, these figures ignore the welfare gains that arose directly

from enjoying better health.

To reconcile the absence of improvements in prevalence and the substantial decline in

self-diagnosed incidence, we go back to the numerous earlier field trials of ITN efficacy and

to accepted epidemiological models of malaria transmission. Consistent with our results,

29

we find that studies that reported both malaria prevalence and incidence systematically

found larger improvements using the latter measures, and some studies (like ours) actually

found increases in prevalence after the introduction of ITNs in treatment areas. Substantive

declines in prevalence have been argued to be unlikely to emerge without large declines

in the number of infective mosquito bites. In turn, such declines are only likely to arise

when ITN coverage is high and/or ITN usage is very regular. By design, our study only

targeted BISWA-affiliated households, with the consequence that even with free distribution

ITN coverage rates rarely surpassed 30% within the village. In addition, although a large

majority of ITNs were retained by households, usage was relatively low at the time of the

follow-up visits, with 45% of free ITNs reported as having being used the night before and

(somewhat surprisingly) lower usage rates for nets purchased on credit.

Even taking the reduction in self-diagnosed incidence at face value, the lack of improve-

ments in RDT-assessed prevalence remain a concern, because it suggests that the potential

reservoir for infection remained unabated in study areas. This is important, because although

asymptomatic cases may not lead to substantive health or economic costs for the individual

affected, sub-clinical malaria can still be transmitted to others, so that it remains “a major

hurdle for malaria elimination, as infected hosts serve as silent reservoirs” (Laishram et al.

2012, p. 9, see also Vinetz and Gilman 2002). Unfortunately, our study was not designed to

measure entomologic indicators such as anopheline density, biting rates and behaviors. We

were thus not able to assess directly any program impacts on these key channels through

which ITNs exert their protective effect.

We also emphasize that low coverage and irregular usage are likely to mimic more closely

the result of actual public health interventions (‘effectiveness’) than studies carried out under

ideal trial conditions (‘efficacy’). From this perspective, the results of our trial should also

be of relevance for the public health literature, given that almost all the results surveyed in

Lengeler (2004) are “from randomized controlled trials where the intervention was deployed

under highly controlled conditions, leading to high coverage and use rates. [...] While

the difference between efficacy and effectiveness is likely to be small for certain medical

interventions (such as vaccination or surgery), it can potentially be large for preventive

interventions such as ITNs” (p. 10). Importantly, the unique features of our study design

also imply that our results should not be interpreted as contradicting such earlier studies on

the efficacy of ITNs. Our findings on the health impacts of ITNs should rather be seen as

complementing the existing literature and suggest that public health interventions which only

achieve the distribution of a relatively limited number of ITNs and/or do not ensure regular

30

usage may fail to achieve the desired effects. Much more may be needed, and efforts should

include ensuring high village-wide coverage, providing incentives for regular use, and possibly

adding complementary interventions such as indoor residual spraying, case management and

environmental measures.

We conclude by emphasizing a number of additional factors that may limit the external

validity of our results. First, although our study area comprised 141 villages from a very

wide geographical area, the study population was not a representative sample of the five

districts where we operated. Our study villages were selected because BISWA already had

a presence there, and only BISWA clients were eligible for the intervention. Therefore, our

study does not identify the impacts of introducing sales of ITNs on credit in a population

with no access to BISWA’s credit network. Extending sales to non-BISWA clients within

our study communities could have increased the overall coverage of ITNs within the village,

but our data are silent about this, and the implications on the expected repayment rates

are unclear. Second, ITN adoption among sample households may have been raised beyond

what achievable by micro-loans only by the information content of the information campaign,

survey, and blood tests that preceded the sales. Should a micro-loan program such as ours

be scaled up, at least some of these factors would be unlikely to be replicated. On the one

hand, we do find evidence that the results of the blood tests (immediately divulged to the

individuals or the guardian) affected demand for ITN offered on credit, and we also observe

higher demand among households included in our baseline survey relative to non-sample

households. On the other hand, we have argued that exposure to the information campaign

was not a plausible confounder, and that being part of the baseline survey can explain at

most part of the success of the micro-loan program at increasing ITN adoption.

One additional limitation is that the bulk of the study was conducted with standard

ITNs, and not with the long-lasting insecticidal nets (LLINs) that are being increasingly

used in many mass distribution campaigns, particularly in Africa. We chose the inclusion

of ITNs for the study because LLINs were not available in the area—and to the best of our

knowledge remain so—and we favored a product that was available locally in case local NGOs

wanted to implement similar interventions. We only adopted LLINs when we implemented

cash sales, after the conclusion of the main study, but funding constraints did not allow us

to evaluate health impacts. The choice of LLINs for free distribution or sale on credit may

have provided a more reliable insecticide concentration on the ITNs in the field, given that

they are factory pre-treated, more wash resistant, and do not need to be re-treated every

six months. In the paper we argue that the guidelines followed for the re-treatment of the

31

standard ITNs adopted in the main study, as well as the choice of insecticide, should have

guaranteed their effectiveness. Despite this, we cannot exclude the possibility that LLINs

may have led to better health impacts.

References

Agha, S., R. Van Rossem, G. Stallworthy, and T. Kusanthan (2007). The impact of a hybrid socialmarketing intervention on inequities in access, ownership and use of insecticide-treated nets.Malaria Journal 6 (1), 13.

Alam, K. and S. Ahmed (2010). Cost recovery of NGO primary health care facilities: a case studyin Bangladesh. Cost Effectiveness and Resource Allocation 8 (1), 12.

Ashraf, N., J. Berry, and J. Shapiro (2010). Can higher prices stimulate product use? Evidencefrom a field experiment in Zambia. American Economic Review 100 (5), 2383–2413.

Ashraf, N., D. Karlan, and W. Yin (2006). Tying Odysseus to the mast: evidence from a commit-ment savings product in the Philippines. Quarterly Journal of Economics 121 (2), 635–672.

Barreca, A. (2010). The long-term economic impact of in utero and postnatal exposure to malaria.Journal of Human Resources 45 (4), 865–892.

Beier, J., G. Killeen, and J. Githure (1999). Short report: entomologic inoculation rates andplasmodium falciparum malaria prevalence in africa. The American Journal of Tropical Medicineand Hygiene 61 (1), 109–113.

Binka, F., F. Indome, and T. Smith (1998). Impact of spatial distribution of permethrin-impregnated bed nets on child mortality in rural northern Ghana. The American Journal ofTropical Medicine and Hygiene 59 (1), 80–85.

Bleakley, H. (2010). Malaria eradication in the Americas: A retrospective analysis of childhoodexposure. American Economic Journal: Applied Economics 2 (2), 1–45.

Cohen, J. and P. Dupas (2010). Free distribution or cost-sharing? Evidence from a randomizedmalaria prevention experiment. Quarterly Journal of Economics 125 (1), 1–45.

Cutler, D., W. Fung, M. Kremer, M. Singhal, and T. Vogl (2010). Early-life malaria exposureand adult outcomes: Evidence from malaria eradication in India. American Economic Journal:Applied Economics 2 (2), 72–94.

D’Alessandro, U., B. O. Olaleye, W. McGuire, P. Langerock, S. Bennett, M. K. Aikins, M. C.Thomson, M. K. Cham, B. A. Cham, and B. M. Greenwood (1995). Mortality and morbidityfrom malaria in Gambian children after introduction of an impregnated bednet programme.Lancet 345 (8948), 479–83.

Das, J., J. Hammer, and C. Sanchez-Paramo (2012). The impact of recall periods on reportedmorbidity and health seeking behavior. Journal of Development Economics 98 (1), 76–88.

de Benoist, B., E. McLean, I. Egli, and M. Cogswell (2008). Worldwide prevalence of anaemia1993-2005 : WHO global database on anaemia. Technical report, World Health Organization.

Devoto, F., E. Duflo, P. Dupas, W. Pariente, and V. Pons (2012). Happiness on tap: Piped wateradoption in urban Morocco. American Economic Journal: Economic Policy 4 (4), 68–99.

Dupas, P. (2012a). Health behavior in developing countries. Annual Review of Economics 3,425–449.

Dupas, P. (2012b). Short-run subsidies and long-run adoption of new health products: Evidencefrom a field experiment. Working Paper.

32

Fink, G. and F. Masiye (2012). Assessing the impact of scaling-up bednet coverage through agricul-tural loan programmes: evidence from a cluster randomised controlled trial in Katete, Zambia.Transactions of The Royal Society of Tropical Medicine and Hygiene 106 (11), 660–667.

Gallup, J. L. and J. D. Sachs (2001). The economic burden of malaria. American Journal ofTropical Medical Hygiene 64 (1 Suppl.), 85–96.

Gersovitz, M. and J. Hammer (2004). The economical control of infectious diseases. EconomicJournal 114, 1–27.

Gersovitz, M. and J. Hammer (2005). Tax/subsidy policies towards vector-born infectious diseases.Journal of Public Economics 89, 647–674.

Hammer, J. (1997). Economic analysis for health projects. The World Bank Research Ob-server 12 (1), 47.

Hausman, J., J. Abreveya, and F. Scott-Morton (1998). Misclassification of the dependent variablein a discrete response setting. Journal of Econometrics 87, 239–269.

Hawley, W. A., P. A. Phillips-Howard, F. O. ter Kuile, D. J. Terlouw, J. M. Vulule, M. Ombok, B. L.Nahlen, J. E. Gimnig, S. K. Kariuki, M. S. Kolczak, and A. W. Hightower (2003). Community-wide effects of permethrin-treated bed nets on child mortality and malaria morbidity in westernKenya. The American Journal of Tropical Medicine and Hygiene 68 (4 Suppl), 121–7.

Hong, S. C. (2007a). The burden of early exposure to malaria in the United States, 1850-1860:Malnutrition and immune disorders. The Journal of Economic History 67 (4), 1001–1035.

Hong, S. C. (2007b). A longitudinal analysis of the burden of malaria on health and economicproductivity: The American case. University of Chicago mimeo.

Killeen, G. F., T. A. Smith, H. M. Ferguson, H. Mshinda, S. Abdulla, C. Lengeler, and S. P.Kachur (2007). Preventing childhood malaria in Africa by protecting adults from mosquitoeswith insecticide-treated nets. PLoS Medicine 4 (7).

Kitchens, C. (2012). A dam problem: TVA’s fight against malaria 1926-1951. Journal of EconomicHistory Forthcoming.

Korenromp, E. (2005). Malaria incidence estimates at country level for the year 2004. Proposedestimates and draft report. World Health Organization: Geneva, Switzerland.

Kremer, M., , E. Miguel, S. Mullainathan, C. Null, and A. P. Zwane (2009). Making Water Safe:Price, Persuasion, Peers, Promoters, or Product Design? Working Paper.

Kremer, M. and E. Miguel (2007). The illusion of sustainability. Quarterly Journal of Eco-nomics 122 (3), 1007–1065.

Kroeger, A., M. Gonzalez, and J. O. nez Gonzalez (1999). Insecticide-treated materials for malariacontrol in Latin America: to use or not to use? Transactions of the Royal Society of TropicalMedicine and Hygiene 93 (6), 565 – 570.

Kumar, A., N. Valecha, T. Jain, and A. P. Dash (2007). Burden of malaria in India: Retrospectiveand prospective view. The American Journal of Tropical Medicine and Hygiene 77 (Suppl 6),69–78.

Laishram, D., P. Sutton, N. Nanda, V. Sharma, R. Sobti, J. Carlton, and H. Joshi (2012). Thecomplexities of malaria disease manifestations with a focus on asymptomatic malaria. MalariaJournal 11 (1), 29.

Lengeler, C. (2004). Insecticide-treated bed nets and curtains for preventing malaria. Cochranedatabase of systematic reviews (Online). Issue 2. Art. No.: CD000363.

Lucas, A. (2010). Malaria eradication and educational attainment: Evidence from Paraguay andSri Lanka. American Economic Journal: Applied Economics 2 (2), 46–71.

33

Malaney, P., A. Spielman, and J. Sachs (2004). The malaria gap. The American Journal of TropicalMedicine and Hygiene 71 (Suppl. 2), 141–146.

McMorrow, M. L., M. Aidoo, and S. P. Kachur (2011). Malaria rapid diagnostic tests in eliminationsettings—can they find the last parasite? Clinical Microbiology and Infection 17 (11), 1624–1631.

Ross, A., N. Maire, L. Molineaux, and T. Smith (2006). An epidemiologic model of severe morbidityand mortality caused by Plasmodium falciparum. The American Journal of Tropical Medicineand Hygiene 75 (Suppl 2), 63–73.

Sachs, J. and P. Malaney (2002). The economic and social cost of malaria. Nature 415, 680–685.

Sahu, S. S., P. Jambulingam, T. Vijayakumar, S. Subramanian, and M. Kalyanasundaram (2003).Impact of alphacypermethrin treated bed nets on malaria in villages of Malkangiri district, Orissa,India. Acta Tropica 89 (1), 55–66.

Sarriot, E. G., E. A. Swedberg, and J. G. Ricca (2011). Pro-sustainability choices and child deathsaverted: from project experience to investment strategy. Health Policy and Planning .

Sharma, S. K., P. K. Tyagi, K. Padhan, A. K. Upadhyay, M. A. Haque, N. Nanda, H. Joshi,S. Biswas, T. Adak, B. S. Das, V. S. Chauhan, C. E. Chitnis, and S. K. Subbarao (2006).Epidemiology of malaria transmission in forest and plain ecotype villages in Sundargarh Districtand Orissa, India. Transactions of the Royal Society of Tropical Medicine and Hygiene 100 (10),917–925.

Smith, M. (2010). The importance of sustainability in international public health NGOs. In P. A.Gaist (Ed.), Igniting the Power of Community, pp. 25–37. Springer New York.

Snow, R. W., C. A. Guerra, A. M. Noor, H. Y. Myint, and S. I. Hay (2005). The global distributionof clinical episodes of Plasmodium falciparum malaria. Nature 434, 214–217.

Strauss, J. and D. Thomas (1998). Health, nutrition, and economic development. Journal ofEconomic Literature 36 (2), 766–817.

Tarozzi, A., A. Mahajan, B. Blackburn, D. Kopf, L. Krishnan, and J. Yoong (2011). Micro-loans,bednets and malaria: Evidence from a randomized controlled trial in Orissa (India). WorkingPaper, available at http://www.econ.upf.edu/~tarozzi/TarozziEtAl2011RCT.pdf.

ter Kuile, F. O., D. J. Terlouw, S. K. Kariuki, P. A. Phillips-Howard, L. B. Mirel, W. A. Hawley,J. F. Friedman, Y. P. Shi, M. S. Kolczak, A. A. Lal, J. M. Vulule, and B. L. Nahlen (2003).Impact of permethrin-treated bed nets on malaria, anemia, and growth in infants in an areaof intense perennial malaria transmission in western Kenya. The American Journal of TropicalMedicine and Hygiene 68 (4 Suppl), 68–77.

Thomas, D., E. Frankenberg, J. Friedman, J.-P. Habicht, M. Hakimi, N. Ingwersen, Jaswadi,N. Jones, K. McKelvey, G. Pelto, B. Sikoki, T. Seeman, J. P. Smith, C. Sumantri, W. Suriastini,and S. Wilopo (2006). Causal effect of health on labor market outcomes: Experimental evidence.Working Paper.

Vinetz, J. M. and R. H. Gilman (2002). Asymptomatic Plasmodium parasitemia and the ecology ofmalaria transmission. The American Journal of Tropical Medicine and Hygiene 66 (6), 639–640.

White, N. J. (2009). Malaria. In G. C. Cook and A. I. Zumla (Eds.), Manson’s Tropical Diseases(22 ed.)., Chapter 73, pp. 1201–1300. London, United Kingdom: Saunders Elsevier.

World Health Organization (2001). Report of the fifth WHOPES working group meeting. Reviewof Olyset Net and Bifenthrin 10% WP, WHO/HQ, Geneva, 3031 October 2001. Technical report,World Health Organization, Geneva, Switzerland. WHO/CDS/WHOPES/2001.4.

World Health Organization (2007). WHO Global Malaria Programme: Position Statement onITNs. Technical report, World Health Organization.

34

http://www.econ.upf.edu/~tarozzi/TarozziEtAl2011RCT.pdf

Tab

le1:

Bas

elin

eSum

mar

ySta

tist

ics

and

Ran

dom

izat

ion

Tes

ts

(1)

(2)

(3)

(4)

(5)

(6)

Con

trol

Fre

eM

Fp-v

alue

s.dev

.O

bs.

1.Sch

edule

dC

aste

/Tri

be/

Oth

erB

ackw

ard

Cas

tes

0.9

(0.0

13)

0.93

3(0

.013

)0.

912

(0.0

21)

0.42

10.

256

1838

2.H

ouse

hol

dsi

ze5.

5(0

.103

)5.

6(0

.117

)5.

3(0

.086

)0.

138

2.22

1844

3.N

o.ch

ildre

nU

5in

hou

sehol

d0.

499

(0.0

33)

0.50

6(0

.030

)0.

487

(0.0

26)

0.89

20.

704

1844

4.M

ale

hou

sehol

dhea

d0.

952

(0.0

09)

0.94

1(0

.011

)0.

932

(0.0

10)

0.36

80.

235

1843

5.H

ouse

hol

dhea

dhas

som

esc

hool

ing

0.72

(0.0

18)

0.70

6(0

.027

)0.

714

(0.0

21)

0.90

80.

452

1843

6.H

ouse

hol

dhea

dhas

seco

ndar

yed

uca

tion

orab

ove

0.08

4(0

.016

)0.

075

(0.0

13)

0.11

4(0

.015

)0.

123

0.28

718

097.

Exp

endit

ure

per

hea

d(2

007

Rs

per

day

)†22

.3(0

.928

)21

.2(0

.827

)24

.2(1

.101

)0.

085*

16.2

1844

8.P

oor

(exp

endit

ure

per

hea

d<

offici

alp

over

tyline)†‡

0.19

5(0

.025

)0.

24(0

.031

)0.

196

(0.0

24)

0.46

30.

408

1844

9.D

ifficu

lt/i

mp

ossi

ble

for

hou

sehol

dto

bor

row

Rs

500

0.52

9(0

.026

)0.

536

(0.0

29)

0.52

9(0

.025

)0.

980

0.49

918

4210

.R

atio

Deb

t/to

tal

year

lyex

pen

dit

ure

0.47

(0.0

82)

0.38

9(0

.048

)0.

400

(0.0

40)

0.68

51.

0118

4311

.H

ouse

hol

dhas

atle

ast

one

net

0.65

4(0

.030

)0.

628

(0.0

29)

0.68

(0.0

23)

0.37

30.

476

1844

12.

Net

s(p

erca

pit

a)†

0.28

7(0

.020

)0.

264

(0.0

18)

0.31

1(0

.018

)0.

167

0.3

1836

13.

ITN

s(p

erca

pit

a)†

0.02

1(0

.006

)0.

046

(0.0

13)

0.05

5(0

.014

)0.

027*

*0.

146

1831

14.

Exp

endit

ure

for

self

-dia

gnos

edm

alar

iala

st6

mon

ths?

565

(77)

725

(72)

686

(85)

0.29

816

8918

4415

.U

sed

net

last

nig

ht†

0.13

1(0

.022

)0.

116

(0.0

19)

0.16

2(0

.017

)0.

195

0.29

518

4416

.U

sed

ITN

last

nig

ht†

0.01

9(0

.006

)0.

022

(0.0

07)

0.03

(0.0

10)

0.61

70.

134

1840

17.

Use

regu

larl

ynet

sduri

ng

“mos

quit

ose

ason

”†0.

564

(0.0

32)

0.51

2(0

.030

)0.

572

(0.0

28)

0.30

40.

453

1844

18.

Pri

cepai

dfo

rb

ednet

s(2

007

Rs)

+82

.9(8

.2)

83.5

(8.4

)72

.9(6

.6)

0.51

063

579

19.

Mal

aria

pre

vale

nce

(RD

T)

0.10

8(0

.016

)0.

115

(0.0

18)

0.12

3(0

.018

)0.

838

0.32

2557

20.

Hem

oglo

bin

(RD

T)

11.0

(0.0

87)

10.7

(0.0

96)

11.0

(0.0

87)

0.13

21.

9125

2821

.A

nem

iapre

vale

nce

(Hb<

11g/

dl)

(RD

T)

0.52

7(0

.024

)0.

569

(0.0

25)

0.50

4(0

.020

)0.

121

0.49

925

2822

.Sel

f-dia

gnos

edm

alar

iaep

isodes

last

6m

onth

s0.

093

(0.0

09)

0.12

4(0

.012

)0.

125

(0.0

12)

0.04

5**

0.32

810

062

23.

Sel

f-dia

gnos

edm

alar

ia/f

ever

epis

odes

last

6m

onth

s0.

218

(0.0

15)

0.23

8(0

.015

)0.

258

(0.0

17)

0.19

60.

446

1006

2

Sou

rce:

Data

from

1844

hou

seh

old

sin

clu

ded

inth

ep

re-i

nte

rven

tion

hou

seh

old

surv

ey(A

pri

l-M

ay

2007).

For

each

vari

ab

le,

colu

mn

s1-3

show

the

exp

erim

enta

larm

-sp

ecifi

cm

ean

san

dth

eco

rres

pon

din

gst

an

dard

erro

rs(i

np

are

nth

esis

)ad

just

edfo

rin

tra-v

illa

ge

corr

elati

on

.C

olu

mn

4re

port

sp

-valu

esfo

ra

test

of

the

nu

llhyp

oth

esis

that

the

mea

ns

are

iden

tica

lacr

oss

the

thre

eex

per

imen

tal

arm

s,w

ith

ast

eris

ks

den

oti

ng

sign

ifica

nce

at

the

10

(*),

5(*

*)

an

d1%

(***)

level

.C

olu

mn

5co

nta

ins

the

stan

dard

dev

iati

on

of

the

vari

ab

leca

lcu

late

dover

the

wh

ole

sam

ple

an

dco

lum

n6

ind

icate

sth

enu

mb

erof

non

-mis

sin

gob

serv

ati

on

s.T

he

un

itof

ob

serv

ati

on

of

the

vari

ab

les

isth

eh

ou

seh

old

inro

ws

1-1

7,

ab

edn

etin

row

18

an

dan

ind

ivid

ual

inro

ws

19-2

3.

Th

ere

sult

sin

row

s19-2

1in

clu

de

on

lyin

form

ati

on

from

ind

ivid

uals

for

wh

om

RD

Ts

wer

eco

nd

uct

ed.

Th

em

ean

sfo

rvari

ab

les

den

ote

d†

wer

ew

eighte

dby

hou

seh

old

size

.+

Mea

nb

edn

etp

rice

sare

esti

mate

das

arm

-sp

ecifi

cm

ean

sof

pri

ces

paid

for

bed

net

sow

ned

by

hou

seh

old

sat

base

lin

e,im

pu

tin

ga

zero

ifth

en

eth

ad

bee

nre

ceiv

edfr

eeof

charg

e.?

For

each

mala

ria

epis

od

e,w

en

ote

dall

the

rela

ted

mon

etary

cost

sas

wel

las

the

nu

mb

erof

days

of

work

or

sch

ool

lost

.H

ealt

hex

pen

dit

ure

sw

ere

elic

ited

usi

ng

an

item

ized

list

that

incl

ud

edd

oct

or

fees

,d

rugs

an

dte

sts,

hosp

italiza

tion

,su

rger

y,co

sts

of

lod

gin

gan

dtr

an

sport

ati

on

(in

clu

din

gth

ose

for

any

care

taker

),lo

stea

rnin

gs

from

days

of

lost

work

,an

dco

stof

non

-hou

seh

old

mem

ber

sh

ired

tore

pla

ceth

esi

ckat

work

.‡“

Poor”

isa

du

mm

yeq

ual

toon

eif

per

cap

ita

month

lyh

ou

seh

old

exp

end

itu

reis

bel

ow

ap

over

tylin

eeq

ual

toR

s381

=326×

(373/319.5

),w

her

e326

isth

eoffi

cial

pover

tylin

efo

rru

ral

Ori

ssa

in2004-0

5,

an

d373

an

d319.5

are

the

Con

sum

erP

rice

Ind

exfo

rA

gri

cult

ura

lL

ab

ore

rsin

May-J

un

e2007

an

dJu

ly2004-J

un

e2005

resp

ecti

vel

y.A

ccord

ing

toth

e2005

Inte

rnati

on

al

Com

pari

son

Gro

up

Glo

bal

Rep

ort

,th

ep

urc

hasi

ng

pow

erp

ari

tyex

chan

ge

rate

was

Rs

14.6

7p

er1

US

D,

seehttp://siteresources.worldbank.org/ICPINT/Resources/icp-final-tables.pdf.

35

http://siteresources.worldbank.org/ICPINT/Resources/icp-final-tables.pdf

Tab

le2:

Bed

net

Acq

uis

itio

nan

dO

wner

ship

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

Inte

rventi

on

(Fall

2007)

Follow

-up

(Win

ter

08-0

9)

ITN

sA

ny

ITN

Bed

net

sSle

pt

Sle

pt

Sle

pt

Sle

pt

Fra

ctio

nD

epen

den

tva

riab

ledel

iver

edD

eliv

ered

own

edu

nd

erund

erun

der

un

-un

der

ofB

ISW

Ab

ednet

ITN

trea

ted

net

BIS

WA

net

ITN

su

sed

(ob

serv

ed)

Fre

e(β

Free

)2.

650.

961.

560.

380.

46-0

.08

0.47

0.45

(0.0

7)(0

.02)

(0.1

09)

(0.0

36)

(0.0

31)

(0.0

26)

(0.0

30)

(0.0

21)

MF

(βMF

)1.

190.

520.

570.

090.

13-0

.04

0.13

0.31

(0.2

1)(0

.05)

(0.1

06)

(0.0

34)

(0.0

26)

(0.0

26)

(0.0

22)

(0.0

29)

Inte

rcep

t(C

ontr

ol)

(β0)

0.30

0.05

0.00

30.

050.

002

(0.0

72)

(0.0

19)

(0.0

07)

(0.0

16)

(0.0

02)

Diff

eren

ce:βFree−βMF

1.46

0.43

0.99

0.29

0.33

-0.0

50.

340.

14p-v

alue

(H0

:βFree−βMF

=0)

0.00

000.

0000

0.00

000.

0000

0.00

000.

1046

0.00

000.

0002

Unit

ofan

alysi

sH

ouse

hol

dH

ouse

hol

dH

ouse

hol

dIn

div

idual

Ind

ivid

ual

Ind

ivid

ual

Ind

ivid

ual

Hou

seh

old

Reg

ress

ion

typ

eL

evel

Lev

elD

DD

DD

DD

DL

evel

Lev

el

Ob

serv

atio

ns

1199

1199

1759

7707

7647

7647

8018

891

R-s

qu

ared

0.55

0.81

0.12

0.09

10.

199

0.00

70.

241

0.03

no.

clu

ster

s94

9414

114

114

114

114

189

Bas

elin

em

ean

ofd

epen

den

tva

riab

leN

/AN

/A1.

580.

130.

020.

11N

/AN

/A

Note

s:S

tan

dard

erro

rs(i

np

are

nth

esis

)are

rob

ust

toin

tra-v

illa

ge

corr

elati

on

.A

llre

gre

ssio

ns

esti

mate

dw

ith

OL

S.

Inco

lum

ns

1,

2an

d8

the

esti

mate

dm

od

elisy

=βM

FMF

+βFreeFree

+u

,w

her

ey

isth

ed

epen

den

tvari

ab

le,

an

dMF

an

dFree

are

du

mm

ies

for

the

two

exp

erim

enta

larm

s.T

hes

ere

gre

ssio

ns

use

on

lyob

serv

ati

on

sfr

om

Fre

ean

dM

Fvilla

ges

.In

colu

mn

s3

to7,

the

mod

elisy

=β0

+βM

FMF

+βFreeFree

+u

,so

thatβM

Fan

dβFree

are

diff

eren

ces

rela

tive

toco

ntr

ols

.T

he

row

lab

eled

‘reg

ress

ion

typ

e’in

dic

ate

sw

het

hery

isth

ele

vel

of

the

vari

ab

lein

dic

ate

din

the

colu

mn

hea

der

or

its

chan

ge

bet

wee

nb

ase

lin

ean

dfo

llow

-up

surv

ey.

Inco

lum

n7,

the

dep

end

ent

vari

ab

leis

ad

um

my

equ

al

toon

ew

hen

an

ind

ivid

ual

was

rep

ort

edas

havin

gu

sed

atr

eate

dn

etth

en

ight

bef

ore

the

inte

rvie

w,

an

dw

hen

the

net

had

bee

nob

serv

edby

the

surv

eyor

an

did

enti

fied

as

on

ed

istr

ibu

ted

thro

ugh

ou

rp

rogra

m.

Inco

lum

n8,

the

dep

end

ent

vari

ab

leis

the

hou

seh

old

-sp

ecifi

cra

tio

bet

wee

nth

enu

mb

erof

BIS

WA

ITN

sin

use

the

nig

ht

bef

ore

the

surv

eyan

dth

enu

mb

erof

BIS

WA

ITN

sd

eliv

ered

toth

eh

ou

seh

old

du

rin

gth

ein

terv

enti

on

.T

his

last

regre

ssio

nis

thu

son

lyes

tim

ate

din

clu

din

gh

ou

seh

old

sth

at

rece

ived

at

least

on

eIT

N.

36

37

Table 3: Correlates of ITN purchase

Dependent variable: at least one ITN purchased

Log(monthly total expenditure per head) -0.116 (0.053)**Debt towards BISWA (per head, quartic root) -0.005 (0.009)Cost of malaria episodes last 6 months (per capita, quartic root)1 0.019 (0.011)*% members who slept under net last night 0.209 (0.093)**% members who slept under ITN last night -0.053 (0.279)# nets owned by household 0.007 (0.026)# nets treated last 6 months -0.033 (0.036)% members using nets during peak season -0.035 (0.079)Any malaria-related deaths last 5 yrs 0.101 (0.141)Expected cost of a malaria episode (quartic root)2 0.014 (0.019)% tested members positive for malaria 0.202 (0.080)**% members with self-reported malaria episodes last 6 months 0.272 (0.116)**Subjective P (malaria | untreated net)− P (malaria | ITN)3 -0.066 (0.106)Subjective P (malaria | no net)− P (malaria | ITN)3 -0.140 (0.142)

Observations 513R-squared 0.11

Notes: OLS estimates of a linear probability model with a binary dependent variable = 1 if the householdpurchased at least one ITN in fall 2007. Standard errors in parenthesis are robust to intra-village correlation.Statistical significance is indicated with * (10% level), ** (5%) and *** (1%). The regressors were measuredat baseline (spring 2007). Only panel households from MF villages are included. Sample size is smallerthan the 589 panel households in MF villages because 76 observations (13%) have at least one regressormissing. Also included in the model are the following regressors, none of which is significant at standardlevels: intercept, age, gender and schooling of household head, household size, number of members youngerthan 5 years old, or 5 to 14, or older than 60, measures of risk aversion and intertemporal preferences. Toreduce the influence of outliers among regressors measured in Rupees, values are transformed into logarithmsor, when zeros are present, using the quartic root, which has a shape similar to the logarithm for positivenumbers (Thomas et al. 2006). Risk aversion is measured by an indicator equal to one when the respondentchose a no-risk lottery from a list of different lotteries (played with real monetary payoff), differing in theexpected value and variance of the reward. We evaluated time preferences with 12 questions where therespondent had to choose between an earlier reward and a later but larger one. The regression includesa dummy equal to one when the respondent always chose the earlier reward, and a variable recording thenumber of “preference reversals” implicit in the choices, which arise when an individual chose a reward atdate t over a larger one at date t+ s but preferred the later reward when the two dates were shifted by anequal time period.1 Includes all actual expenses for in-patient and out-patient care, drugs, transportation and lost householdearnings. 2 Expected total cost of a malaria episode for a working adult male, including all items listed above.3 The probabilities were elicited by asking respondents to express the likelihood of an event by choosing aninteger between zero (impossible event) and ten (certainty).

38

Table 4: Results of Cash InterventionVillages # LLINs Any LLIN

A. All villages 40 0.159 0.108(0.033) (0.019)

B. Earlier Control villages (PC) 25 0.162 0.103(0.043) (0.022)

C. New villages 15 0.152 0.119(0.048) (0.037)

D. Low price 20 0.226 0.149(0.069) (0.039)

E. High price 20 0.100 0.073(0.026) (0.019)

F. Baseline households 25 0.199 0.121(0.061) (0.026)

G. Non-baseline households 25 0.151 0.098(0.046) (0.025)

Tests (p-values)H0: B=C 0.8779 0.7203H0: D=E 0.0944 0.0863H0: F=G 0.4638 0.4501

Source: authors’ calculations from 2012 data from Cash villages. All standard errors and tests are robustto intra-village correlation of residuals. The 25 PC villages are a subset of the 47 Control villages initiallyincluded in the study, while the 15 New villages had not been selected before. ‘Sample households’ arehouseholds in PC villages that had been earlier selected as sample households for the 2007-2009 study. Thelower prices were Rs 200 for a single LLIN and 250 for a double, while the higher prices were respectivelyRs 240 and 300. The price paid by our research team to Sumitomo, the manufacturer, was about twice aslarge.

Tab

le5:

Impac

tof

Inte

rven

tion

onR

DT

-bas

edH

ealt

hIn

dic

es

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

+ve

Mala

ria

Rob

ust

ness

Ch

eck

s+

ve

Mala

ria

Hem

oglo

bin

An

em

ic(H

b<

11g/d

l)T

este

rF

EIR

SD

um

mie

s

Fol

low

-up

DD

Fol

low

-up

DD

Fol

low

-up

DD

Fol

low

-up

DD

Fol

low

-up

DD

Fre

edis

trib

uti

on=

10.

037

0.05

4-0

.033

0.22

20.

01-0

.024

0.02

10.

038

0.04

0.06

2[0

.030

][0

.040

][0

.105

][0

.107

]**

[0.0

22]

[0.0

33]

[0.0

26]

[0.0

36]

[0.0

35]

[0.0

39]

Mic

ro-l

oans=

10.

044

0.06

30.

023

0.04

60.

005

0.03

50.

023

0.04

60.

035

0.05

5[0

.035

][0

.039

][0

.094

][0

.123

][0

.021

][0

.035

][0

.029

][0

.036

][0

.030

][0

.040

]C

onst

ant

0.18

30.

063

11.4

330.

277

0.38

4-0

.111

0.37

90.

227

0.18

50.

064

[0.0

22]*

**[0

.028

]**

[0.0

64]*

**[0

.075

]***

[0.0

12]*

**[0

.024

]***

[0.0

43]*

**[0

.047

]***

[0.0

25]*

**[0

.031

]**

Dum

mie

sfo

rSpra

yin

gN

oN

oN

oN

oN

oN

oN

oN

oY

esY

esT

este

rF

ixed

Eff

ects

No

No

No

No

No

No

Yes

Yes

No

No

Obse

rvat

ions

7154

1897

7149

1869

7149

1869

7154

1897

7154

1897

No.

clust

ers

(villa

ges)

141

141

141

141

141

141

141

141

141

141

R-s

quar

ed0.

0022

0.00

370.

0001

0.00

360.

0001

0.00

210.

0467

0.04

150.

0051

0.00

41

Fre

e=M

F(p

-val

ue)

0.83

30.

8289

0.60

580.

1568

0.84

740.

0937

*0.

9502

0.82

000.

8893

0.85

84F

ree=

MF

=0

(p-v

alue)

0.35

380.

228

0.87

490.

1025

0.90

430.

2437

0.64

790.

3971

0.38

990.

2407

Not

es:

Dat

afr

omb

asel

ine

(Sp

rin

g20

07)

and

pos

t-in

terv

enti

on

hou

seh

old

surv

eys

(Win

ter

2008-0

9).

All

resu

lts

are

OL

Ses

tim

ate

sw

ith

ind

ivid

ual-

leve

lob

serv

atio

ns.

Sta

nd

ard

erro

rs(i

nb

rack

ets)

are

robu

stto

intr

a-v

illa

ge

corr

elati

on

.A

ster

isks

ind

icate

sign

ifica

nce

at

the

10

(*),

5(*

*)

an

d1%

(***)

leve

l.E

stim

ates

inco

lum

ns

2,4,

6,8

and

10(D

D)

on

lyin

clu

de

test

sfr

om

ind

ivid

uals

test

edb

oth

at

base

lin

ean

dat

follow

-up

.In

colu

mn

s7

an

d8

we

incl

ud

efi

xed

effec

tsfo

rth

ein

div

idu

als

wh

oca

rrie

dou

tth

eb

lood

test

sd

uri

ng

the

stu

dy.

Inco

lum

ns

9an

d10

regre

ssors

als

oin

clu

de

du

mm

ies

for

inn

erw

alls

hav

ing

bee

nsp

raye

din

2008

/09,

asi

mil

ar

du

mm

yfo

rsp

rayin

gof

ou

ter

wall

san

dtw

od

um

mie

s=

1w

hen

info

rmati

on

ab

ou

tsp

rayin

gis

mis

sin

gfo

rin

ner

orou

ter

wal

lsre

spec

tive

ly.

39

Tab

le6:

Impac

tof

Inte

rven

tion

onSel

f-re

por

ted

Mal

aria

Indic

es

(1)

(2)

(3)

(4)

(5)

(6)

(7)

Mal

aria

Mal

aria

and

epis

od

esin

6m

onth

sb

efor

ein

terv

iew

pre

vio

us

Nu

mb

erof

Day

sof

wor

kH

ealt

hex

pen

dit

ure

s#

epis

od

es#

epis

od

esp

aid

mon

thep

isod

esor

school

lost

All

Doct

ors

pai

dfo

rfo

rw

ith

low

er&

dru

gsw

ith

deb

tco

nsu

mp

tion

Fre

ed

istr

ibu

tion

=1

-0.0

04-0

.048

-1.9

-194

-86

-0.1

1-0

.12

[0.0

04]

[0.0

18]*

*[1

.2]

[180

.1]

[100

.1]

[0.0

5]**

[0.0

6]**

Mic

ro-l

oan

s=1

-0.0

02-0

.051

-2.4

-269

.3-1

87.2

-0.1

1-0

.09

[0.0

04]

[0.0

18]*

**[1

.1]*

*[1

43.4

]*[7

6.5]

**[0

.05]

**[0

.05]

*

Con

stan

t(C

ontr

ol)

-0.0

010.

025

1.5

238.

216

8.8

0.07

-0.1

8[0

.003

][0

.013

]*[0

.8]*

[110

.2]*

*[5

4.6]

***

[0.0

4]*

[0.0

3]**

*

En

dli

ne

leve

l(C

ontr

ol)

0.00

70.

115

5.8

862.

848

6.8

0.22

0.07

Un

itof

obse

rvat

ion

Ind

ivid

ual

Ind

ivid

ual

Hou

seh

old

Hou

seh

old

Hou

seh

old

Hou

seh

old

Hou

seh

old

Ob

serv

atio

ns

8684

8684

1768

1768

1768

1768

1768

Fre

e=M

F=

0(p

-val

ue)

0.60

650.

0087

***

0.08

62*

0.17

130.

053*

0.03

26**

0.07

7*F

ree=

MF

(p-v

alu

e)0.

5873

0.85

570.

7043

0.65

780.

3105

0.94

20.

6285

Not

es:

Dat

afr

omb

asel

ine

(Sp

rin

g20

07)

and

post

-inte

rven

tion

hou

seh

old

surv

eys

(Win

ter

2008-0

9).

All

resu

lts

are

OL

Ses

tim

ate

sof

diff

eren

ce-

in-d

iffer

ence

sm

od

els.

All

outc

omes

refe

rto

mala

ria

epis

od

esd

iagn

ose

das

such

by

the

resp

on

den

t.M

on

etary

valu

esare

in2008-0

9R

up

ees

an

dar

eat

the

hou

seh

old

leve

l.In

colu

mn

4,“a

ll”

hea

lth

exp

end

iture

sw

ere

elic

ited

usi

ng

an

item

ized

list

that

incl

ud

edd

oct

or

fees

,d

rugs

an

dte

sts,

hos

pit

aliz

atio

n,

surg

ery,

cost

sof

lod

ging

and

tran

sport

ati

on

(in

clu

din

gth

ose

for

any

care

take

r),

lost

earn

ings

from

day

sof

lost

work

,an

dco

stof

non

-hou

seh

old

mem

ber

sh

ired

tore

pla

ceth

esi

ckat

work

.S

tan

dard

erro

rs(i

nb

rack

ets)

an

dte

sts

are

rob

ust

toin

tra-v

illa

ge

corr

elati

on

.A

ster

isks

ind

icat

esi

gnifi

can

ceat

the

10(*

),5

(**)

and

1%(*

**)

leve

l.

40

A Appendix - NOT FOR PUBLICATION

A.1 Selection of Sample Villages and their Representativenesswithin the Study Districts

The villages included in our sample were selected from a list of 878 villages where BISWAoperated in 2007. These villages were spread across 318 panchayats (administrative unionsof villages) in 26 blocks across the five districts of Bargarh, Balangir, Keonjhar, Khandhamaland Sambalpur (see Figure A.1). We selected 150 villages for the study, stratified as follows:33 from Balangir, 48 from Bargarh, 30 from Keonjhar, 9 from Khandhamal and 30 fromSambalpur (the allocation was approximately proportional to the number of BISWA com-munities in each district). Villages were drawn using a pseudo-random number generator,with a selection algorithm that ensured the inclusion of a multiple of three villages from eachblock. Blocks where the Government of Orissa was planning to initiate free distribution ofnets were excluded from the sampling frame. While the study locations were thus chosento minimize this risk, the sampling scheme was designed to preserve the balanced structureof the sample across treatment groups in case the state Government initiated any unantic-ipated distribution. Data collected during the post-intervention survey show that indeeddistribution of nets from the Government (or from other NGOs) was extremely limited instudy areas, see Tables 2 and A.12. After the baseline survey, but before the intervention,nine of the 150 villages were found to have no actual BISWA activity and were then excludedfrom the study. Data from these villages are excluded from the analysis.

In Table A.7, we evaluate the characteristics of communities in our sample relative toother communities in the five study districts, by using data from the 2001 Census of Indiaon a broad range of village-level characteristics. Overall, the five study districts included apopulation of 8,991 villages. Although the data used in this paper have been collected from2007 onwards, the time gap relative to the 2001 census is short enough that a comparisonbetween sample and non-sample villages should be informative.

The results show that the null hypothesis of equality of means between sample andnon-sample villages is strongly rejected for most village characteristics (column 6). Samplevillages are relatively large (both in terms of area and population), with mean total popu-lation more than twice as large as in non-sample villages. Sample villages also appear to becloser to towns, although not to a large extent. Mean distance from the nearest town is 35kilometers among non-sample villages and 1-10 kilometers less in sample villages. Ameni-ties are overall significantly better in sample villages as reflected, for instance, in the higherproportion of villages with schools, health centers, a post office, a telephone connection andelectricity. Interestingly, sample villages are also characterized by significantly larger frac-tions of land devoted to rice cultivation. This may have implication on malaria prevalence,because rice fields are often an ideal breeding ground for larvae of Anopheles mosquitoes.

We also test the null hypothesis that village characteristics are on average equal in thethree experimental arms (column 7). This is useful, because the randomization tests inTable 1 only evaluated balance in household-level characteristics among villages included atbaseline. In a list of 26 variables, the test of equality across groups is only rejected, at the10% level, for the presence of a middle school in the village.

A.1

A.2 Details of Blood Tests

The RDTs were conducted using fingerprick samples of less than 0.5 ml of blood for eachtest. Malaria prevalence was determined using the Binax Now malaria RDT. This testis well validated in comparison to blood smears for the diagnosis of malaria. The RDTdetects both current and recent infections, up to 2-4 weeks prior to the test. The resultof the RDT is read on a test strip, located on a card, where a reagent is added to theblood sample. Recent infection is detected when the presence of Plasmodium antigens inthe blood (histidine-rich protein 2, or HRP2) is signaled by the appearance of darker lineson the white strip. High concurrency between test readers (including non-trained ones) hasbeen documented in clinical trials of the RDT (Khairnar et al. 2009). The test does notindicate the level of parasitemia, and only delivers a positive/negative result for malariainfection, besides showing whether that infection is due to P. falciparum, to one of the otherPlasmodium species, or to both (Moody 2002, Farcas et al. 2003, van den Broek et al. 2006,Khairnar et al. 2009). The test has been shown to have both good specificity and sensitivity.Both these concepts are defined assuming that the “null hypothesis” of the test is that theindividual does not have malaria. The specificity is calculated as the fraction of negativecases correctly diagnosed as such (that is, it is equal to one minus the probability of a Type-Ierror). The sensitivity is the fraction of positive cases correctly diagnosed as such (that is,one minus the probability of a Type-II error).

Hemoglobin levels were tested with the HemoCue 201 Hb analyzer, a portable, accuratesystem for measuring Hb. The test, like the one used to detect malaria prevalence, requiresless than 0.5 ml of blood and delivered results in approximately 15 minutes.

A.3 Gender and Age variation in Malaria and Anemia Rates atBaseline

In Figure A.2 we show malaria and anemia prevalence by gender and age group. Womenwere 3 pp more likely to test positive for the parasite, and the difference is significant atthe 5% level. There is overall little variation in prevalence by age group, although when wedisaggregate the data into single-year age bins we find that prevalence follows an invertedU-shape pattern with respect to age (results not shown).28 Such age patterns are commonlyobserved in malarious areas, because very young children have initially some immune pro-tection from the mother (and are more often protected by bednets, when available) althoughsuch immunity is gradually lost and subsequently replaced by their own semi-immunity ac-quired through repeated exposure to the disease, so that malaria prevalence usually peaks forchildren of age 2-10 years (see Smith et al. 2007 for a review of the evidence). Sharma et al.(2006) found similar non-monotone age gradients in incidence and prevalence in Sundargarh,a district of Orissa that shares borders with two of our study districts.

There was substantial variation in anemia rates by gender and age. Approximately 80%of tested U5, of either gender, were anemic. Anemia rates declined significantly amongadults aged 15 to 45, but prevalence remained extremely high (60%) among women, whileit was less than 12% among men. Prevalence increased again among older adults, where it

28The same inverted U-shaped patterns was also found at follow-up, results not shown.

A.2

characterized about three-quarters of women and one quarter of men. Similar patterns foranemia for different ages and genders are common in developing countries (see for instanceThomas et al. 2006), and are also present in data from Orissa collected as part of the IndianNational Family and Health Survey in 2004-05, which showed an anemia prevalence of 65%among U5, 34% among women 15-49 and only 8% among men in the same age group.

A.4 Details of Bednets and Treatment with Insecticide

The nets were of uniform quality, composed of white polyester multifilament, mesh size 156,and 75 denier. They had a bottom reinforcement of 28 cm, with single nets measuring180×150×100 cm and double nets measuring 180×150×160 cm. A total of 6,750 single and3,250 double nets were supplied by Biotech International Limited, who generously donated5,000 single and 2,500 double nets.

The bednets were treated on the spot at the time of delivery by trained personnel,following rules recommended in World Health Organization 2002, using K-Othrine flow,which contains deltamethrin, a highly effective pyrethroid. The subsequent re-treatmentsafter six and 12 months were done similarly by our trained collaborators, using the sameguidelines. While wearing gloves, the field worker dipped the washed net into a bucket wherewater had been mixed with the appropriate quantity of insecticide. After being soaked fora few minutes, the net was removed from the bucket and was laid flat on a plastic sheet ormat in the shade to dry. The concentration of the insecticide was determined based on themanufacturer’s instructions: 10 ml of insecticide to 500 ml of water for single nets and 15ml-750 ml for double nets. The chemical concentration made re-treatment optimal after sixmonths.

The study design did not incorporate the systematic use of ‘bioassays’, that is, proceduresto test rigorously the insecticidal power of treated bednets. However, at the conclusionof the study, samples from four ITNs gathered from Free villages were tested through gaschromatographic analysis, and two of the nets still had concentrations of deltamethrin aroundthe concentrations recommended by the WHO (15-25 mg/m2), while the other two bednetshad lower concentrations. Although the number of ITNs tested is obviously very small, theresults do not signal obvious shortcomings with the re-treatment operations, given that thebednets had been last re-treated 6-7 months earlier, and that it is not unexpected to find lowinsecticide concentrations six months after re-treatment (particularly if the ITN has beenwashed multiple times).

Pyrethroids have been widely used for bednet impregnation with encouraging evidenceabout the lack of side-effects on human health (World Health Organization 2005). In Orissa,synthetic pyrethroids have been in use since 1999, and tests performed in 2002-03 in severaldistricts (including our study districts Balangir, Khandhamal and Keonjhar) showed highrates of susceptibility to deltamethrin of Anopheles culicifacies and A. fluviatilis, the twomost common malaria vectors in the state (Sharma et al. 2004). The insecticidal efficacyof deltamethrin compound has also been confirmed in Sundargarh, which borders the studydistrict Sambalpur (Yadav et al. 2001, Sharma et al. 2006).

A.3

A.5 Attrition

At follow-up, the survey team attempted to re-contact all households included in the base-line survey. The survey protocol called for at least three attempts, although a handful ofhouseholds were re-contacted after 4 or 5 visits. Refusals accounted for only 13 of 76 losthouseholds. As a result, attrition was limited, and of the 1,844 initial households, 1,768(96%) were re-interviewed. Attrition was 5% in MF and control villages and 3% in Freecommunities (see Table A.8, column 2). The null of equal attrition rates among arms is notrejected at standard levels, regardless of whether we use individual or joint tests. There waslittle correlation between attrition and household characteristics at baseline, including RDTresults and bednet ownership and usage (columns 3 and 4). The only regression coefficientsthat are individually statistically significant indicate that households with an older and bet-ter educated head were less likely to exit the panel. On the other hand, we cannot reject thejoint null that all the included slopes are equal to zero (p-value=0.14).

We also investigated whether significant changes in household composition took placebetween the baseline and the follow-up survey, as well as whether such changes were bal-anced across experimental arms. This was potentially important for two reasons. First,changes in availability of ITNs may have arisen from changes in the number and age ofhouseholds members (for instance, young children often share a sleeping space with theirparents). Second, malaria and anemia prevalence at baseline differed across age and gendergroups (see Figure A.2), so that changes in the demographic structure of the household mayhave confounded aggregate changes in such health measures calculated over all householdmembers. We looked at both entry into or exit from panel households and to changes in therelative weight of different demographic groups. This analysis was possible because enumer-ators filled a complete household roster both at baseline and at follow-up, so that we canseparately identify new members as well as individuals who left the household because ofdeath or relocation. We find that these factors did not plausibly drive any of the results inthe paper. We omit the detailed analysis for brevity but the interested reader can find it inthe appendix of Tarozzi et al. (2011).

A.6 The Information Campaign and Household Survey as PossibleConfounders

In principle, the relatively high ITN adoption rates observed with micro-loans may havebeen explained at least in part by the information campaign (IC) and household surveycum RDTs that preceded the sales. These factors may have made the malaria problemmore salient, leading to high demand regardless of the possibility to pay over time ratherthan in cash. There is indeed growing awareness within field-based development economicsthat surveys may themselves constitute ‘interventions’, see e.g. Zwane et al. (2011). In thissection we argue that although such confounders likely played a role, they cannot plausiblyexplain more than a fraction of the high demand observed with micro-loans.

As a first point, we note that confounders were also present in the recent seminal studiesthat documented very steep demand curves among poor populations in developing countries.ITN sales in Cohen and Dupas (2010) took place at ante-natal visits, during which theimportance of ITNs was discussed and hemoglobin levels were measured (p. 14). In Ashraf

A.4

et al. (2010), the baseline survey also included a number of questions on water use practicesand Clorin adoption, as well as measurements of the concentration of chlorine in households’drinking water supply (pp. 2389-2391). In this experiment, the water disinfectant Clorin wassold during door-to-door marketing visits. The de-worming project studied in Kremer andMiguel (2007) was carried out with teacher training, teacher and NGO-led school lessons,and a number of classroom educational materials (pp 1013-1015). In addition, in that studythe huge drop in demand for drugs observed after the introduction of cost-sharing was alsoobserved in areas where pupils had been tested for intestinal worms infections and had beenpart of the de-worming campaign. From this perspective, our study design was comparableto that of these earlier studies and we show that, if anything, it could perhaps be singledout for its unusual ability to study the impact of such behavioral components.

A.6.1 The IC as a possible confounder

We first discuss the IC, which we argue was not a plausible key confounder. First, the ICwas a simple one-time presentation about malaria, the means by which it is transmitted andthe importance and rationale for ITN use, a demonstration of how to hang nets properly,and advice on proper use and re-treatment. Such presentation usually lasted less than onehour, and a large majority of households were already familiar with the IC content, with themajor exception of the importance of treating bednets regularly. For instance, at the time ofthe baseline survey, 96% of respondents stated (un-prompted) that malaria was transmittedby mosquitoes, while 95% stated that bednets can prevent the disease (although less than3% explicitly mentioned ‘ITNs’ rather than ‘bednets’). Second, the IC conducted before thesales on credit in 2007 and the one before the cash sales in 2011 were very similar, and yet theresulting demand was significantly different. Third, we demonstrated that in control areasthere was virtually no change in ITN usage between baseline and follow-up (Table 2, column5). During the same period there was only a small increase (0.3 bednets per households) inthe number of bednets owned, suggesting that the IC did not change behavior or perceptionsof malaria risk substantively. Fourth, additional evidence comes from a household surveyconducted at the same time as the follow-up survey—in Winter of 2008-09—in 25 villagesthat had not been part of the initial study.

These 25 villages were added specifically to allow for the separate identification of anyimpact of the IC and/or of the survey itself on behavior. These 25 ‘follow-up only’ villages(‘FUO’ hereafter) were selected from the same randomly sorted lists used for the selection ofthe communities at baseline. In other words, we did not complete a new randomization, butwe selected the “next 25 villages” from the same randomization done in 2007. The similarityof the new village relative to those included since baseline was confirmed by comparing thevillage characteristics included in Table A.7 (measured during the 2001 Census) between the25 FUO villages and the 141 study villages where the baseline survey had been conducted.The null of equality is rejected for only three of the 26 characteristics (results available uponrequest). In each FUO village, 15 households were selected regardless of BISWA affiliation,using simple random sampling from publicly available census lists formed in 2002 as part ofthe ‘Below Poverty Line census’ by the Government of Orissa. Because BISWA had a strongpresence in the study areas, the sample ended up including BISWA households in almost allvillages (21/25).

A.5

When we compare sample households in Control areas to BISWA households in FUOvillages, we find that the number of bednets was very close between the two groups, andthe null of equality cannot be rejected at standard levels: the mean was 0.36 per personin Control and 0.32 in FUO, and the p-value for the test of equality is 0.3248. Consistentwith this result, the survey-elicited subjective probability of someone falling ill with malariawithin a year when always sleeping under an ITN was 0.16 in both sub-samples.29 Overall,then, the data do not support the hypothesis that the IC affected behavior or perceptionsabout malaria substantively.

A.6.2 The baseline survey and RDTs as possible confounders

The baseline survey included a long list of questions about malaria and bednets. In addition,the results of the RDTs were available on the spot, a few minutes after the blood samplewas taken, and individuals were immediately informed about the outcome of the test. Thesefactors may have made the disease more salient, possibly increasing the willingness to payfor ITNs regardless of the possibility being offered to delay payment. Indeed we have shownthat demand was significantly higher among households where at least one member testedpositive to the blood test. We argue, however, that these factors cannot plausibly explainmore than a fraction of relatively high demand for ITNs on credit when compared withearlier studies that found very little demand for health-protecting technologies when thesewere not offered for free.

First, we have discussed before how comparable confounders (including health tests) werealso present in earlier studies that found very low demand for health products. In principle,such confounders may have been more important in our empirical context, but it is not clearwhy this should be the case.

Second, we have shown that, by comparing outcomes in Control areas with those ofBISWA households in FUO villages, we found no evidence that the joint impact of the IC,the baseline survey and RDTs increased bednet ownership or changed perceptions about theeffectiveness of ITNs. Even so, we cannot rule out the possibility that demand would havebeen higher in the New villages in the Cash arm if we had filled the same questionnaireand conducted the same RDTs in these communities (these elements could not be addedto the supplemental arm due to time and funding constraints). In PC villages, however,both potential confounders had been present, albeit more than four years prior to the Cashintervention. As we pointed out earlier, there is no difference in ITN adoption between PCand New villages which at least suggests that the surveys and RDTs had no longer termeffects on take-up. In addition, within PC villages, demand is very similar (and low) whenwe directly compare households who had been exposed to the survey and RDTs, and otherswho had not (see rows F and G of Table 4). Recall also that attrition between baseline andfollow-up was very limited, so almost all sample households in PC villages had been exposedto a lengthy questionnaire and RDTs both at baseline, in 2007, and at follow-up, in 2008-09.

To probe this issue further, we can use data about ITN purchases in MF villages amongBISWA households not included in the pre-intervention survey, among whom biomarkers

29These subjective probabilities were elicited by asking respondents to place a number of marbles rangingfrom 0 to 10 into a cup, with the number increasing in the probability of the event taking place in the future.Similar methodologies have been adopted in several studies, see Delavande et al. (2010) for a review.

A.6

were not collected. At the time of the MF sales, in 2007, surveyors recorded the number andtype of ITNs purchased by all BISWA members, regardless of their inclusion in our sample.Our data do not include the total number of ‘BISWA households’ in study villages, but thisnumber can be estimated from the lists of BISWA members supplied by the micro-lenderat the beginning of the study. The latter figure is not the correct one to be used for theestimation of demand among non-sample households, for two reasons. First, some householdshad more than one member affiliated to BISWA (on average 1.11). Second, a fraction ofindividuals listed as BISWA members were found not to be such during the field work, or hadmigrated, or were otherwise excluded from the study population. In this way, we estimatethat every 100 members listed by BISWA corresponded to about 79 BISWA households. Letn and ns denote respectively the total number of buyers in MF villages and the number ofbuyers among sample households. Let also m denote the initial number of BISWA membersprovided by the micro-lender, and let ms denote the number of baseline sample householdsin the same villages. We thus calculate demand among non-sample BISWA households as(n−ns)/(0.79m−ms) = 0.28. Uptake was then about twice as large as that observed amongBISWA members who were offered LLINs for cash at the same nominal price (.149, see Table4), and about four times as large as that observed when the price was kept constant in realterms (.073). In addition, as described earlier, these figures likely attenuate the differences indemand between Cash and MF, because the voucher system implies that a BISWA memberwho was not present during the voucher distribution would not be counted in the demandestimation, rather than being counted as not having purchased.

Another key factor points to the fact that the 28% take up rate among ‘non-sample’BISWA households in MF communities is artificially biased downwards relative to demandamong sample households. That is, field reports indicate that more effort was put intoensuring attendance of sales meetings for sample relative to non-sample households. In fact,during the first sale session, 78% of baseline households attended the sale, while only 56% didamong non-baseline households. Similarly, during the second session, conducted 1-2 weeksafterwards, attendance rates were 62 and 40% for the former and latter group respectively.Of course, attendance itself may have been influenced by the inclusion in the baseline.

To summarize: we argue that while the IC and the baseline surveys may have played arole in increasing take-up, the effects are not sufficient to explain the overall take-up rates.

A.7 Respondent-reported Malaria Incidence versus RDT Results

As we mention in Section 4, our data on malaria incidence are derived from respondentreports and not from blood tests. Such reports may be noisy indicators of actual incidenceand may also suffer from bias potentially differential across experimental arms. For instance,the distribution of ITNs may have made the disease more salient, pushing respondents toover-report illnesses or it may have led to a decrease in the perceived malaria risk, withopposite effects on program impacts. In this section we provide evidence in support of theview that, despite these concerns, incidence data in our data set were a valuable source ofinformation on malaria burden.

First, note that reported incidence can be validated against the RDTs only for very recentmalaria episodes, because the RDTs we used in the field can only detect malaria episodesthat are still ongoing or that took place no more than 2-4 weeks earlier (see Appendix A.2).

A.7

Let the binary variable Si = 1 if individual i was reported as having had malaria in the monthpreceding the survey, and let Mi = 1 if the individual tested positive for malaria when testedwith a RDT. In our post-intervention sample, there is a total of 63 individuals for whomSi = 1 and for whom we observe Mi. Among these 63 individuals, 28 (44%, 95% C.I. 0.32-0.57) also have Mi = 1. As we discuss in the paper, most malaria cases detected by theRDTs were apparently asymptomatic and thus not mentioned by respondents, but despitethis the self-reported information about recent malaria incidence is strongly correlated withthe RDTs. To show this we estimate with OLS the following model, using all individuals forwhich Mi and Si are non-missing

Si = β0 + βMMi + ui.

The estimated intercept is β0 = 0.006 while βM = 0.012 and is significant at the 1% level(p-value= 0.006, adjusted for clustering at the village level, n = 7, 153). In other words,self-reported recent incidence was three times as large for individuals who tested positiverelative to others who did not.

That respondents were able to recognize symptomatic malaria episodes is also confirmedby the fact that the results are very different if we estimate a regression such as the one aboveusing as dependent variable a dummy = 1 if the individual was only reported as having had‘fever’ during the last month. In this case, the intercept is 0.03 while βM = 0.002 and is notsignificant at any standard level (p-value= 0.715, adjusted for clustering at the village level,n = 7, 153).

Another key observation is that the link between Si and Mi does not appear to bedifferential across experimental arms, so there is no compelling evidence that the interventionchanged perceptions about malaria incidence conditional on actual malaria infection. Thefractions P (Mi = 1 | Si = 1) are 42% in Control areas (8/19), 45% in Free (10/22) and 45%in MF (10/22). The fractions are thus almost identical, and the null of equality cannot berejected (p-value= 0.9724 for the joint null of equality. The individual differences are alsonot significant).

A.8 Post-intervention RDT Success Rates

In the post-intervention survey, all members of households re-contacted after the baselinewere targeted for blood tests. Our testers were able to successfully test 75% of members inpanel households, while 19% could not be tested because they were not present at the timeof the visits and only 6% because consent was not given, see columns 1 and 4 in Table A.10.The figures in columns 2 and 5 show that absence and refusal were almost identical acrossexperimental arms. Conversely, we find differences in testing success across different agegroups (columns 3 and 6). Almost one third of adult males (15-45, the omitted category inthe regressions) could not be tested because of absence during the visits, probably becausethey were more likely to be off to work. Testing rates among all other demographic groupswere substantively and statistically significantly higher, especially among U5 of either genderand among women 15 years old and above. For these groups, testing rates were close to 90%.The testing rates are very close between boys and girls, and the null of equality betweengenders cannot be rejected for both U5s and 5 to 15 year old children. Refusal rates were

A.8

highest among women over 45 (8%) and girls U5 (9%). Refusal rates were 3 pp lower amongU5 boys relative to girls but the null of equality between genders cannot be rejected atstandard significance levels.

A.9 Changes in Malaria Indices by Demographic Group

Was the lack of health benefits shared by all demographic groups? The bars in Figure A.3show malaria and anemia prevalence for each experimental arm by gender and age group,together with 95% confidence intervals.

Among adult males (age 15 or above), malaria prevalence was ∼15% and almost identicalacross arms (panel A). Among U5s, prevalence was 11% in control villages but about twiceas large in intervention communities: 18.4% in Free and 19.8% in MF villages. However, theestimates are imprecise, and the difference relative to control is not significant at standardlevels, although the p-values are relatively small (below 0.2). Details of the test statisticsare available upon request from the authors. Prevalence among males is highest among 5-14boys, where in each arm it is ∼15 pp higher than for younger children, so that the differencesamong groups are almost identical in these two age groups.

These patterns change when we look at females (panel B), although again differencesbetween arms are never significant at standard levels. Among females, we observe almostidentical prevalence across arms among the youngest girls (∼15%) and higher prevalence inintervention villages in older age groups. In each experimental arm, the highest prevalenceis observed among females of age 5 to 59.

Overall, these results document remarkable differences in malaria prevalence across sub-groups, but these differences are largely concentrated between genders or across age groupsrather than across experimental arms. Note also that, consistent with the baseline results,we do not observe prevalence rates monotonically declining with age. The relatively lowprevalence among U5s is actually driven by very low rates among children less than twoyears old (results not in the figure). Of a total of 263 children in this latter age group, only12 (4.6%) tested positive, while prevalence jumps to 23.3% among the 412 two to four yearsold tested. Overall, in our sample malaria prevalence peaks among 5 to 10 years old, andthen gradually declines with age. These patterns are similar among experimental arms.

Consistent with the baseline results, the results for anemia (panels C and D) show largesystematic gaps across gender-age groups. In particular, these results confirm the U-shapeof anemia prevalence with respect to age for both genders, as well as the significantly higheranemia rates among females 5 and older relative to males of the same age. Like for malaria,however, the differences in anemia prevalence between arms are small and never significantat standard levels.

A.10 RDT Validation Study

In July 2009, we carried out a small validation study after the conclusion of the follow-upsurvey in collaboration with the Malaria Research Centre (MRC) Field Station in Rourkela(Orissa), which confirmed the accuracy of the RDTs. A total of 205 blood samples wereindependently collected from the MRC team from individuals with malaria symptoms fromthree villages. The RDT cards were interpreted by three different blinded readers, including

A.9

two of the testers who were part of the field team during our study, and the most seniorsurvey monitor in our research team. These results were then compared with thick and thinblood smears read with microscopy by the MRC team for the same samples, with the smearresult accepted as the correct infection status. The results showed very high sensitivity(> 90% for each of the three readers, see Table A.11 for details). The fraction of correctlyidentified negatives (specificity) ranged from 74 to 85%.

The lower specificity (higher prevalence) measured by the RDTs relative to microscopywas not surprising, given that these tests may detect the presence of the P. falciparumantigens up to 2-4 weeks after parasitemia has cleared (Humar et al. 1997). The RDT resultswere overall very similar but not identical between readers (pairwise correlations ranged from0.78 to 0.88). In columns 9 and 10 of Table 5, we show that the ITT estimates for malariaprevalence remain almost identical if we include tester fixed effects in the regressions.

A.11 Changes in Other Prophylactic Behavior

In Table A.12, we look at differences among experimental arms in knowledge about causesof malaria (panel A), precautions one can take against it (panel B) and wall spraying be-tween baseline and follow-up (panel C). The survey instrument asked respondents—withoutprompting—to list all possible causes of malaria, and then asked “[w]hat are the best pre-cautions you can take to protect yourself from getting malaria.” In each arm, 85% or moreof respondents list mosquito bites as a cause of malaria. Overall, households in interventioncommunities appear to be about as knowledgeable regarding causes of malaria as those incontrol areas, although the test of equality is rejected at the 10% level (but not at the 5%) forthree of the four causes of malaria, and in each of these cases it is one of the experimental armsthat suggests the best knowledge. There was no systematic variation in malaria-avoidingbehavior among groups (panel B). Bednets are by far the most commonly listed precaution,mentioned by 82-87% of respondents (with the highest proportions in intervention villages).The next two most common precautions are “avoid contaminated environment” (16-21%)and “avoid drinking contaminated water” (5-8%). For all the fourteen indices, the test ofequal means is not rejected at the 5% level, although the null is rejected at the 10% in twocases, and the joint null of equality for all behaviors is rejected (p-value = 0.0421). However,the differences are not consistent with risk-averting behavior being more common in controlvillages, and indeed in several cases they indicate the opposite (for example, use of smoke orlong sleeves, or cleaning of drainage pools).

In panel C we analyze differences in residual spraying of indoors or outdoor walls. Al-though the null hypothesis of equal proportion among treatment groups cannot be rejectedat standard levels, the magnitude of the differences between control and intervention areasis large. The reason why the null is not rejected despite the large differences is that theintra-village correlation for these two variables is very large (0.41 and 0.63 for inner andouter spraying respectively). Our data do not tell us if these differences were driven byhousehold decisions, or if instead they resulted from choices made by public health officialswho may have scheduled wall spraying taking into account our intervention. To evaluatewhether differences in spraying rates help explain the lack of health benefits in interventionvillages, we re-estimate the ITT model for malaria prevalence including dummies for recentwall spraying among the regressors, but this leaves the estimated impacts almost identical

A.10

(see columns 9 and 10 of Table 5).

A.12 Changes in Local Anopheles Behavior or Resistance to In-secticide

In principle, changes in the characteristics of the local Anopheles population may explainthe lack of improvements in malaria and anemia prevalence. First, Anopheles mosquitoesmay have been resistant to deltamethrin, the insecticide used to impregnate study bednets,or they may have developed resistance during the course of the study. Second, the reductionin malaria transmission may have been hampered if local Anopheles took a sufficiently highfraction of blood meals outside of the sleeping hours, when individuals were less likely tobe protected by ITNs. In principle, a large increase in the fraction of individuals protectedby bednets, as well as the excito-repellent property of deltamethrin, could lead to changesin peak biting hours, or in indoors vs. outdoors feeding habits. The increased difficulty infinding blood meals during the sleeping hours could force mosquitoes to increase biting attimes when individuals are not protected by ITNs. Our project did not collect informationon the local Anopheles population, before or after the intervention, so we cannot addressthese concerns directly. However, a number of factors make these hypotheses unlikely tohold.

First, recent studies carried out in Orissa suggest that local Anopheles biting patternsand susceptibility to deltamethrin made ITNs a promising protective tool against malaria. InKeonjhar, one of our study districts, Sahu et al. (2009) found that biting activity of the mainlocal malaria vectors was concentrated between 2100 and 0300 hours, regardless of the season.Sharma et al. (2004) describes tests performed in 2002-03 in several Orissa districts (includingour study districts Balangir, Khandhamal and Keonjhar). The tests showed high ratesof susceptibility to deltamethrin of Anopheles culicifacies and A. fluviatilis, the two mostcommon malaria vectors in the state. The insecticidal efficacy of deltamethrin compoundhas also been confirmed in Sundargarh, which borders the study district Sambalpur (Yadavet al. 2001, Sharma et al. 2006). The field work for these studies was conducted a fewyears before our project, but a very recent study in Sundargarh, conducted in 2009-2010,found that synthetic pyrethroids were still highly effective against both A culicifacies and Afluviatilis, despite the fact that study areas had been exposed to either large-scale sprayingwith pyrethroids or to large-scale free distribution of bednets treated with deltamethrin, thesame synthetic pyrethroid adopted in our study (Sharma et al. 2012). Another recent study,carried out in 2009 in Madhya Pradesh, central India, found some evidence of resistanceto deltamethrin, but even in areas that had been sprayed regularly in the previous 5-10years, the researchers documented about 75% mortality rates in the local population of Aculicifacies when exposed to the chemical (Mishra et al. 2012).

Second, although the emergence of resistance to insecticides such as DDT and pyrethroidshas been documented following widespread use in agriculture or wall spraying, there is as yetlittle evidence of resistance developing as a consequence of the introduction of ITNs. Even insituations where resistance is present, ITNs have been documented to retain some protectiveefficacy (Enayati and Hemingway 2010). The only exception we are aware of is Trape et al.2011. In this study, the authors found that the introduction of deltamethrin-treated LLINs

A.11

in one village in Senegal led initially to sharp reductions in malaria incidence and prevalence,but that resistance to the insecticide became widespread in about two years. This led to anincrease in malaria morbidity relative to before LLINs distribution among adults and olderchildren. However, unlike in our study, nets were distributed to all villagers, and ownershipand usage rates remained around 60-80% throughout the study period (and were close to100% at the onset of the study).

Third, the literature is overall inconclusive about the impact of ITNs on Anopheles bitingpatterns, with only a fraction of the evidence pointing to changes in mosquito behavior thatmay have reduced the efficacy of nets (Takken 2002, Pates and Curtis 2005). After thedistribution of permethrin-treated bednets to all inhabitants of one hamlet in Papua NewGuinea, Charlwood and Graves (1987) observed a relative increase in biting during theevening, although the number of Anopheles in the area decreased substantially. Similarresults were also found after mass distribution of ITNs in five villages in Tanzania (Magesaet al. 1991) and in locations where ITNs were distributed to cover all beds in Kenya (Mbogoet al. 1996) and Benin (Moiroux et al. 2012). Note that in all these studies ITNs had beendelivered to ensure universal coverage, a situation in stark contrast with our case.

In sum, the existing evidence points to the likely efficacy of deltamethrin-treated ITNsin our study areas, and the literature suggests that the relatively low coverage of ITNs atthe community level would have been unlikely to produce the emergence of either insec-ticide resistance or changes in biting patterns that may have reduced the benefits of theintervention.

A.13 Impacts on Self-reported Incidence Adjusted for Misdiag-noses

In Appendix A.7 we have shown that only 44% of the individuals reported as having hadmalaria in the same month as the interview tested positive to malaria. Some of theseindividuals may have recovered from malaria by the time blood samples were taken, but it islikely that the discrepancy is at least partly explained by misdiagnoses. In malarious areas,while asymptomatic cases are common, it is also common to attribute to malaria other feverepisodes not caused by this disease (see for instance Adhvaryu 2012, Cohen et al. 2012).In such case, the figures in column 2 of Table 6 could confound changes in symptomaticmalaria cases with changes in other symptomatic fever episodes. In addition, our data showthat respondents were also misdiagnosing some malaria episodes as ‘fever’. This can be seenlooking at the RDT results among individuals reported as having had fever during the samemonth as the interview. Among these 221 individuals, we find that 22% tested positive tomalaria (95% C.I. 0.16-0.29).

In this section we use these considerations to construct a procedure to adjust the impactson incidence in column 2 of Table 6 in a way that takes misdiagnosis into account. Notethat we are not interested in estimating the program impacts on ‘true’ malaria incidence(regardless of whether an episode was recognized by the respondent), but rather we aim atestimating impacts on symptomatic malaria incidence. We argue in the paper that the latteris of interest because it measures cases severe enough to be perceived and to lead to illness-related costs recognized by the respondent. Recall that our data include both self-reported

A.12

malaria cases and self-reported fever cases. Suppose that MA,ti represents the number ofmalaria cases during the previous six months reported for individual i from experimental armA (A = Free,MF,Control) interviewed at time t (where t = 0 denotes baseline and t = 1denotes follow-up), while FA,ti is the corresponding figure for self-reported fever incidence.We assume that errors of diagnoses for symptomatic cases happen at the same rate over timeand across different experimental arms (see Appendix A.7 for some evidence in support ofthis assumption). Consistent with the estimates above, we then assume that only 44% ofself-reported malaria cases are actually malaria, but also that 22% of self-reported fever caseswere actually malaria. We can thus estimate the mean number of actual malaria episodesat time t in a given treatment arm as 0.44 × MA,t + 0.22 × FA,t, where MA,t and FA,t arerespectively malaria and fever incidence as measured in our raw recall data.

Next, let βDDY,T denote the estimated difference-in-difference impact on outcome Y = M,Ffor treatment T = MF,Free versus control areas. From column 2 of Table 6 we estimatethat βDDM,Free = −0.048 and βDDM,MF = −0.051. Similarly, when we estimate the same model

using F as dependent variable we obtain (results not shown in the table) βDDF,Free = −0.032

(s.e. 0.033, so not significant) and βDDF,MF = −0.056 (s.e. 0.031, significant at the 10%level). Because the DD is a linear combination of time and arm-specific means, and underthe previously stated assumption that mid-diagnosis errors of symptomatic illness are non-differential across arms and over time, the adjusted DD for actual symptomatic malariaincidence can be finally calculated as 0.44 × βDDM,T + 0.22 × βDDF,T , T = MF,Free. Thefinal estimates are thus −0.048 × 0.44 − 0.032 × 0.22 = −0.028 (s.e. 0.012) in Free and−0.051× 0.44− 0.056× 0.22 = −0.035 (s.e. 0.011) in MF areas.30 Such estimates are thusabout 60% as large as those in column 2 of Table 6, although both remain significant andsubstantive in magnitude.

A.14 Epidemiological Models of ITN Use

Current advances in epidemiological models of malaria transmission may help explaining thelink between ITN usage and coverage and changes in malaria indices in our study areas. Inparticular, Killeen et al. (2007) describe a complex model that describes how malaria infec-tion is affected by several factors, including mosquito numbers, biting patterns and mortality(also in relation to ITN presence) and above all ITN coverage and usage. The model simulatesthe protective power of ITNs for both users and non-users, by calibrating 16 exogenouslydetermined factors (largely borrowed from earlier studies), and then showing how ITN protec-tion varies with changes in coverage and usage. The protective power of bednets is measuredas a relative risk (RR) of entomological inoculation rates (EIR), that is, the number of infec-tive bites per year calculated relative to a situation where no ones uses nets. A useful featureof this study is that the authors also provide a spreadsheet that can be used to analyze howchanges in any of the exogenous factors affect the RR. The spreadsheet is available at http://www.ncbi.nlm.nih.gov/pmc/articles/PMC1904465/bin/pmed.0040229.sd001.xls.

Panel A in Figure A.4 shows one of the key results in Killeen et al. (2007): in a scenariowhere about 60% of the population always uses ITN, individuals without an ITN (the dashed

30To take into account that both 0.22 and 0.44 were estimated, we calculate the standard errors using1,000 block bootstrap replications, using the village as the block.

A.13

http://www.ncbi.nlm.nih.gov/pmc/articles/PMC1904465/bin/pmed.0040229.sd001.xls


line) are as protected as an individual who sleep regularly under an ITN in a communitywhere no one else does (the continuous line). A key assumption to produce these results isthat the index individual using the net does so very regularly. Specifically, panel A (identicalto one of the graphs in Figure 3 of Killeen et al. 2007) is produced assuming that theindividual uses the net for 90% of the potential time of exposure. In our empirical context,we have argued that information on previous night usage of ITNs is reliable, but even so ourdata indicate that only 45% of the program ITNs were in use the night before the follow-up survey in Free villages, and about 30% were in use in MF communities. If we assumethat the frequency of usage is equal to such cross-sectional usage rates, while leaving allother parameters in the epidemiological model unchanged, the association between relativetransmission intensity and coverage for users becomes as described in panel B of Figure A.4.Even under such scenario ITNs would provide some protection, but if less than 20% of thevillage population always uses ITNs (as surely is the case in a large majority of our studyareas), then the RR remains close to 0.6-0.7.

These results formalize the intuition that sleeping under an ITN, even when done ir-regularly, should be expected to decrease the number of infectious bites to some extent.However, whether the decline is sufficient to produce a decline in malaria prevalence (ourkey biomarker) is not obvious. The link between EIR and prevalence is studied in Beieret al. 1999, who analyze data from 31 studies throughout Africa where both outcomescould be estimated. They find that, after excluding two clear outliers the data are tightlyconcentrated around a linear regression relationship between malaria prevalence and thelogarithm of annual EIR (R2 = 0.712). Malaria prevalence predicted by the linear fit is24.68 + 24.2 log10EIR, with the standard errors of intercept and slope equal respectively to3.06 and 5.42.

Although there is no direct information from Indian locations, the authors point out that“[w]hile malaria stratification according to ecologic zones is an important element of malariacontrol, it is important to note that the fundamental relationships between EIR and theprevalence of P. falciparum infection will likely hold across diverse ecosystems in Africa.”Together with the very tight distribution of the scatterplot around the regression line linkingEIR to prevalence (see their Figure 2), this suggests that a similar relationship will also likelyhold outside of the African continent. In areas neighboring our study districts, Sharma et al.(2006) documented EIR in the range of 3-114 infective bites per year, depending on location,well inside the relevant range considered in Beier et al. 1999.

In our study areas, prevalence rates was about 20%, with village-specific prevalenceranging from 0 to about 60% and 95% of the 141 study villages showing prevalence below0.53. Looking at Figure 2 of Beier et al. (1999), this suggests that the EIR in the area waslikely between 1 and 10, but it also suggests that a 30-40% decline in EIR may have barelyaffected prevalence, given that EIR in the 1 to 10 range are associated with a very widespectrum of prevalence rates. Using the words in Beier et al. (1999), “it may not be possibleto achieve dramatic decreases in prevalence of P. falciparum infection at sites in Africa unlesscontrol measures reduce EIRs to levels well below 1 infective bites per year” (p. 111). Ourresults suggest that similar arguments will hold in other malaria-endemic areas outside ofthe African continent, such as our study areas in Orissa.

A.14

A.15 Malaria Prevalence and ITN Coverage

Recall that only BISWA clients received free ITNs or the offer of ITNs for sale on credit.Although BISWA has a large presence in the study area, we estimate that on average only20% of people lived in households with at least one BISWA affiliate and thus were eligible forinclusion in the study.31 It is now accepted that the externalities offered by mass adoptionof ITNs are a key factor for ITN efficacy, although the relative role of personal versus massprotection of ITNs is not yet well understood (Binka et al. 1998, Hawley et al. 2003, Killeenet al. 2007). Reductions in malaria indices have been documented among non-users of ITNsliving within a few hundred meters of communities covered by mass distribution of ITNs.In our intervention, study villages were scattered spatially over a very broad geographicalarea (see Section 2), so cross-village externalities are not plausible. Here, we look at therelationship between village-level coverage and changes in malaria prevalence in our studyarea.32

As a first step, we estimated village-specific changes in malaria prevalence in all interven-tion communities. We then plot the results against a measure of village-wide ITN coverage,calculated as the ratio of the total number of ITNs distributed to BISWA households (re-gardless of their inclusion in the survey sample) and village population counts from the 2001Indian Census. Although not up-to-date, the population counts are a good proxy for currentpopulation, and if anything, in most cases 2001 population would underestimate currentpopulation, so that our estimates may overstate true coverage. The results are displayedseparately for MF and Free communities in the two panels at the top of Figure A.5. Eachgraph also shows the fitted values of two OLS regressions, one where we include data fromall villages (the continuous line) and the other where we exclude the very few villages wherethe ITN coverage ratio was larger that 0.35 (the dashed lines).

When we include all Free villages, there is a positive association between malaria preva-lence at follow-up and program coverage. The estimated slope (0.59) is actually significantat the 1% level. However, the results are driven by the three outlier villages with coverage> 0.35, and when we exclude them the slope becomes negative but very close to zero (−0.02)and not significant at standard levels (p-value = 0.966). In MF villages (panel B), wheresubstantially fewer ITNs were distributed, the slope of the regressions are negative but wecannot reject the null that slopes are zero at standard levels, although when we include allvillages the slope is almost significant at the 10% level (p-value = 0.103).

Because the ITN coverage achieved in MF communities was endogenously determined byhousehold purchase decisions, its association with changes in malaria prevalence should not

31We estimated the fraction by making use of village population data from the 2001 census of India,together with estimates of the total number of individuals living in households with at least one BISWAmember. Let sv and bv denote respectively average household size and average number of BISWA affiliatesin BISWA households in village v, both estimated using baseline survey data. Let also mv be the numberof BISWA members in the village, as provided by the micro-lender. Then, if we denote by pv the villagepopulation from the census, our estimate of the fraction who lives in BISWA households is sv(mv/bv)/pv.

32We could not study the link between prevalence and density of ITNs throughout the village (e.g. thenumber of ITNs per squared hectare), because we do not have information on village size. The Indian Censusreports the area covered by each village, but it does not report the size and the distribution of the areascovered by dwellings. In Section A.16 we look at the link between prevalence and coverage within the villageusing data from a subset of communities where we collected GIS information for all households.

A.15

be interpreted as necessarily causal. In contrast, in communities with free distribution, thenumber of ITNs delivered was decided by our research team based on household size andcomposition. This produced variation in ITN coverage resulting only from the distributionof BISWA affiliation and household composition within the community. Even so, BISWAaffiliation could be associated with village characteristics related to malaria prevalence, al-though if we regress malaria prevalence at baseline on ITN coverage the slope is close to zero(0.03) and not significant (p-value = 0.720). On the one hand, the fact that the dependentvariable in panels A and B is the change in prevalence, eliminates any possible spuriouscorrelation due to time-invariant (observed or unobserved) village-level characteristics. Onthe other hand, there may be other unobserved differences in trends correlated with bothITN coverage and malaria prevalence. To address this concern, in panel C of Figure A.5 welook at the relationship between changes in prevalence and the fraction of the populationaffiliated to BISWA in control villages (“BISWA penetration”). No ITNs were distributedin these communities, but by construction BISWA penetration is very strongly correlatedwith the measure of ITN coverage that would have been observed if ITNs had been dis-tributed as in Free communities. Indeed, the correlation between the two variables in Freevillages is 0.95. The graph in panel C shows no clear association between changes in malariaprevalence and BISWA penetration. This suggests, albeit indirectly, that the lack of anassociation between changes in prevalence and ITN coverage in Free villages (panel A) isunlikely to be caused by differential trends in prevalence across communities with varyingdegrees of BISWA penetration.

As an additional check, we use Control and Free villages to estimate an OLS regressionof the village-level change in prevalence on BISWA penetration, the Free dummy, and theinteraction between the two variables. If ITN coverage were causing declines in malariaprevalence in our sample, we would expect the coefficient on the interaction to be negative.Consistent with the results in panel A, we find instead that the coefficient is positive andsignificant when we include all 94 villages (= 1.8, p-value= 0.009), and close to zero and notsignificant (= 0.25, p-value= 0.770) when we exclude the three villages with coverage largerthan 0.35. Overall, we conclude that in our sample the coverage of the intervention did notappear to be systematically related to the changes in malaria prevalence.

A.16 Within-village Externalities

In Section A.15 we found no direct support for the link between ITN coverage and malariaprevalence. In principle, it is still possible that such a link existed within villages, with moreprotection provided in clusters with a denser concentration of ITNs. Although the baselineand follow-up surveys did not include geo-coding of household locations, such informationwas recorded later in a subset of 11 study villages, including four Control and seven Freevillages. The geo-coding was completed in February-June 2012, during the implementationof the supplemental Cash arm described in Section 3.2.1. Unfortunately, time and fundingconstraints did not allow us to conduct a complete mapping of the whole study area. In thissection we show that the available data provide some evidence of within-village externalities,although the estimates are very imprecise and the null of no effect can never be rejected.

In each of the 11 villages, surveyors visited all households, regardless of BISWA affiliation,

A.16

and recorded for each latitude and longitude using GPS hardware.33 Surveyors also recordedwhether the household belonged to BISWA at the time of the baseline survey, in 2007.Although the GPS survey was carried out a few years later, we were able to find nearly all ofthe original surveyed households and field observations suggested that few households hadmoved within the village so we are reasonably confident that the 2012 GPS coordinates areaccurate measures of households’ 2007-2009 locations.

We then constructed measures of population density within pre-specified radii of our sam-ple households. Concretely, for each sample household (an ‘index’ household) we constructedthe total number of neighbors (P ) and the number of BISWA households (B) within a givenradius. The number of BISWA households matters because they all received ITNs in the Freevillages, so that B provides a good proxy for the potential ITN coverage around the indexhousehold. Controlling for total population in the neighborhood is important, because Bis by construction strongly correlated with population density around the index household,and this in turn may be correlated with unobserved characteristics that could be linked tohealth. On average households had 16 neighbors within a 20-meter radius, of whom 7 wereBISWA members.

We thus estimate the following model for the malaria indicator Miv of individual i invillage v:

Miv = αv + αPPiv + αBBiv + τPPiv × Freev + τBBiv × Freev + εiv,

where αv is a village fixed effect, and Freev is the usual dummy for Free villages. The inclu-sion of Control villages allows to interpret the estimates of τP and τB as causal, because anycorrelation between malaria infection and population density regardless of ITN presence willbe captured by αP and αB. In particular, if there are externalities from being surroundedby households with ITNs, we expect τB < 0, that is, after controlling for total density P , anincrease in the number of BISWA neighbors should be associated with lower malaria preva-lence in Free relative to Control villages. In contrast, we do not have clear predictions forthe sign of τP , which measures the impact of population density regardless of ITN coverage.Note that the interpretation of τB as measuring externalities needs to be taken with caution,given that the number of BISWA neighbors, even when controlling for overall density, wasnot randomly determined, and may proxy for other unobserved location characteristics.34

Overall, we have malaria infection status for 611 individuals, but because identificationrelies on individual variation in neighbors interacted with treatment status, we cluster stan-dard errors at the village level. Because we have only 11 villages, we estimate standarderrors using block bootstrap, using the village as the block in each iteration. We only fo-cus on relatively short radii, because several of the villages are small, and using a radiusof 50 meters or more would generate collinearity between the measures of density and thevillage fixed effects, reducing further the already small number of observations. Table A.15displays the results, which show some evidence of externalities at short ranges, from 5 to

33For each location, two independent measurements were taken, and both were recorded. This doublemeasurement allowed to detect a handful of measurement errors, but otherwise the vast majority of mea-surements were almost identical, so the results remain virtually unchanged if we use either one or the othersets of records.

34So, for instance, our specification is not identical to that in Miguel and Kremer (2004) or Dupas (2012b),because in both these studies the fraction of treated neighbors was randomized by design.

A.17

20 meters, although the estimates are very imprecise and the null of no correlation is neverrejected at standard levels. The estimated τB becomes close to zero for distances of 30 or40 meters, but the point estimates are relatively large if we look at households immediatelyaround the index households. For instance, If we compare two households in Free villageswith an average number of total neighbors within a 10m radius (5.7), but with 0 versus 2.6(the average) BISWA members among them, the predicted probability of malaria will be2.6× 0.035 = 9 percentage points lower in the household with more BISWA neighbors, rela-tive to what would be predicted in Control areas. Of course, the 95% confidence interval islarge, so the null of no relationship, or even of a positive relationship between ITN coverageand malaria cannot be rejected.

Appendix Bibliography

Adhvaryu, A. (2012). Learning, misallocation, and technology adoption: Evidence from new malariatherapy in Tanzania. Working Paper.

Ashraf, N., J. Berry, and J. Shapiro (2010). Can higher prices stimulate product use? Evidencefrom a field experiment in Zambia. American Economic Review 100 (5), 2383–2413.

Beier, J., G. Killeen, and J. Githure (1999). Short report: entomologic inoculation rates andplasmodium falciparum malaria prevalence in africa. The American Journal of Tropical Medicineand Hygiene 61 (1), 109–113.

Binka, F., F. Indome, and T. Smith (1998). Impact of spatial distribution of permethrin-impregnated bed nets on child mortality in rural northern Ghana. The American Journal ofTropical Medicine and Hygiene 59 (1), 80–85.

Charlwood, J. and P. Graves (1987). The effect of permethrin-impregnated bednets on a populationof Anopheles farauti in coastal Papua New Guinea. Medical and veterinary entomology 1 (3),319–327.

Cohen, J. and P. Dupas (2010). Free distribution or cost-sharing? Evidence from a randomizedmalaria prevention experiment. Quarterly Journal of Economics 125 (1), 1–45.

Cohen, J., P. Dupas, and S. G. Schaner (2012). Price subsidies, diagnostic tests, and targetingof malaria treatment: Evidence from a randomized controlled trial. NBER Working Paper No.17943.

Delavande, A., X. Gine, and D. McKenzie (2010). Measuring subjective expectations in developingcountries: A critical review and new evidence. Journal of Development Economics 94 (2), 151–163.

Dupas, P. (2012). Short-run subsidies and long-run adoption of new health products: Evidencefrom a field experiment. Working Paper.

Enayati, A. and J. Hemingway (2010). Malaria management: Past, present, and future. AnnualReview of Entomology 55 (1), 569–591.

Farcas, G. A., K. J. Y. Zhong, F. E. Lovegrove, C. M. Graham, and K. C. Kain (2003). Evaluationof the Binax now(r) ict test versus polymerase chain reaction and microscopy for the detection ofmalaria in returned travelers. The American Journal of Tropical Medicine and Hygiene 69 (6),589–592.

A.18

Hawley, W. A., P. A. Phillips-Howard, F. O. ter Kuile, D. J. Terlouw, J. M. Vulule, M. Ombok, B. L.Nahlen, J. E. Gimnig, S. K. Kariuki, M. S. Kolczak, and A. W. Hightower (2003). Community-wide effects of permethrin-treated bed nets on child mortality and malaria morbidity in westernKenya. The American Journal of Tropical Medicine and Hygiene 68 (4 Suppl), 121–7.

Humar, A., C. Ohrt, M. A. Harrington, D. Pillai, and K. C. Kain (1997). Parasight(R)F testcompared with the polymerase chain reaction and microscopy for the diagnosis of Plasmodiumfalciparum malaria in travelers. The American Journal of Tropical Medicine and Hygiene 56 (1),44–48.

Khairnar, K., D. Martin, R. Lau, F. Ralevski, and D. Pillai (2009). Multiplex real-time quanti-tative PCR, microscopy and rapid diagnostic immuno-chromatographic tests for the detectionof Plasmodium spp: performance, limit of detection analysis and quality assurance. MalariaJournal 8 (1), 284.

Killeen, G. F., T. A. Smith, H. M. Ferguson, H. Mshinda, S. Abdulla, C. Lengeler, and S. P.Kachur (2007). Preventing childhood malaria in Africa by protecting adults from mosquitoeswith insecticide-treated nets. PLoS Medicine 4 (7).

Kremer, M. and E. Miguel (2007). The illusion of sustainability. Quarterly Journal of Eco-nomics 122 (3), 1007–1065.

Magesa, S., T. Wilkes, A. Mnzava, K. Njunwa, J. Myamba, M. Kivuyo, N. Hill, J. Lines, andC. Curtis (1991). Trial of pyrethroid impregnated bednets in an area of Tanzania holoendemicfor malaria. Part 2. Effects on the malaria vector population. Acta Tropica 49, 97–108.

Mbogo, C. N. M., N. M. Baya, A. V. O. Ofulla, J. I. Githure, and R. W. Snow (1996). Theimpact of permethrin-impregnated bednets on malaria vectors of the Kenyan coast. Medical andVeterinary Entomology 10 (3), 251–259.

Miguel, E. and M. Kremer (2004). Worms: identifying impacts on education and health the presenceof treatment externalities. Econometrica 72 (1), 159–217.

Mishra, A. K., S. K. Chand, T. K. Barik, V. K. Dua, and K. Raghavendra (2012). Insecticideresistance status in Anopheles culicifacies in Madhya Pradesh, central India. Journal of VectorBorne Diseases 49 (1), 39–41.

Moiroux, N., M. B. Gomez, C. Pennetier, E. Elanga, A. Djenontin, F. Chandre, I. Djegbe, H. Guis,and V. Corbel (2012). Changes in Anopheles funestus biting behaviour following universal cov-erage of long-lasting insecticidal nets in Benin. Journal of Infectious Diseases.

Moody, A. (2002). Rapid diagnostic tests for malaria parasites. Clinical Microbiology Reviews 15 (1),66–78.

Pates, H. and C. Curtis (2005). Mosquito behavior and vector control. Annual Review of Entomol-ogy 50, 53–70.

Sahu, S. S., K. Gunasekaran, and P. Jambulingam (2009). Bionomics of Anopheles minimus and An.fluviatilis (Diptera: Culicidae) in East-central India, endemic for falciparum malaria: Humanlanding rates, host feeding, and parity. Journal of Medical Entomology 46 (5), 1045–1051.

Sharma, S., A. Upadhyay, B. Kindo, M. Haque, and P. Tyagi (2012). Impact of changing over ofinsecticide from synthetic pyrethroids to DDT for indoor residual spray in a malaria endemicarea of Orissa, India. The Indian Journal of Medical Research 135 (3), 382–388.

Sharma, S. K., P. K. Tyagi, K. Padhan, A. K. Upadhyay, M. A. Haque, N. Nanda, H. Joshi,S. Biswas, T. Adak, B. S. Das, V. S. Chauhan, C. E. Chitnis, and S. K. Subbarao (2006).Epidemiology of malaria transmission in forest and plain ecotype villages in Sundargarh Districtand Orissa, India. Transactions of the Royal Society of Tropical Medicine and Hygiene 100 (10),

A.19

917–925.

Sharma, S. K., A. K. Upadhyay, M. A. Haque, O. P. Singh, T. Adak, and S. K. Subbarao (2004).Insecticide susceptibility status of malaria vectors in some hyperendemic tribal districts of Orissa.Current Science 87 (12), 1722–1726.

Smith, D., C. Guerra, R. Snow, and S. Hay (2007). Standardizing estimates of the Plasmodiumfalciparum parasite rate. Malaria Journal 6 (1), 131.

Takken, W. (2002). Do insecticide-treated bednets have an effect on malaria vectors? TropicalMedicine & International Health 7 (12), 1022–1030.

Tarozzi, A., A. Mahajan, B. Blackburn, D. Kopf, L. Krishnan, and J. Yoong (2011). Micro-loans,bednets and malaria: Evidence from a randomized controlled trial in Orissa (India). WorkingPaper, available at http://www.econ.upf.edu/~tarozzi/TarozziEtAl2011RCT.pdf.

Thomas, D., E. Frankenberg, J. Friedman, J.-P. Habicht, M. Hakimi, N. Ingwersen, Jaswadi,N. Jones, K. McKelvey, G. Pelto, B. Sikoki, T. Seeman, J. P. Smith, C. Sumantri, W. Suriastini,and S. Wilopo (2006). Causal effect of health on labor market outcomes: Experimental evidence.Working Paper.

Trape, J.-F., A. Tall, N. Diagne, O. Ndiath, A. B. Ly, J. Faye, F. Dieye-Ba, C. Roucher,C. Bouganali, A. Badiane, F. D. Sarr, C. Mazenot, A. Toure-Balde, D. Raoult, P. Druilhe,O. Mercereau-Puijalon, C. Rogier, and C. Sokhna (2011). Malaria morbidity and pyrethroid re-sistance after the introduction of insecticide-treated bednets and artemisinin-based combinationtherapies: a longitudinal study. The Lancet Infectious Diseases 11 (12), 925–932.

van den Broek, I., O. Hill, F. Gordillo, B. Angarita, P. Hamade, H. Counihan, and J.-P. Guthmann(2006). Evaluation of three rapid tests for diagnosis of P. falciparum and P. vivax malaria incolombia. The American Journal of Tropical Medicine and Hygiene 75 (6), 1209–1215.

World Health Organization (2002). Instructions for treatment and use of insecticide-treatedmosquito nets. WHO/CDS/RBM 2002.41, World Health Organization, Geneva, Switzerland.

World Health Organization (2005). Safety of pyrethroid for public health use.WHO/CDS/WHOPES/GCDPP/2005.10, WHO/PCS/RA/2005.1, Communicable DiseaseControl, Prevention and Eradication WHO Pesticide Evaluation Scheme (WHOPES) &Protection of the Human Environment Programme on Chemical Safety (PCS).

Yadav, R. S., R. R. Sampath, and V. P. Sharma (2001). Deltamethrin treated bednets for control ofmalaria transmitted by Anopheles culicifacies (Diptera: Culicidae) in India. Journal of MedicalEntomology 38 (5), 613–622.

Zwane, A. P., J. Zinman, E. V. Dusen, W. Pariente, C. Null, E. Miguel, M. Kremer, D. S. Karlan,R. Hornbeck, X. Gine, E. Duflo, F. Devoto, B. Crepon, and A. Banerjee (2011). Being surveyedcan change later behavior and related parameter estimates. Proceedings of the National Academyof Sciences 108 (5), 1821–1826.

A.20

http://www.econ.upf.edu/~tarozzi/TarozziEtAl2011RCT.pdf

Study Design: Location

Malaria: “number one public health problem” in Orissa (OHDR, 2004)

2003 Dept of Health and Family Welfare data show 417,000 cases of malaria (83%falciparum).

High % of self-reported malaria (NFHS-1999) in our study districts: 8.5% inSambalpur and Bargarh, 8.8% in Balangir, 12.3% in Keonjhar, 17.2% in Phulbani.

Tarozzi et al. (Duke, RAND and Stanford) Micro-loans, Bednets and Malaria April 2009 1 / 1

Figure A.1: Study AreasNotes: Study communities at baseline included 30 villages in Sambalpur, 9 in Khandhamal, 30 in Keonjhar(Kendujhar), 33 in Balangir and 48 in Bargarh. Nine villages were later excluded from the analysis becausethe baseline survey showed that BISWA had no active presence there (5 villages in Sambalpur and 4 inBalangir).

0.2

.4.6

.8

U5 15 to 45 45 or olderFemale Male Female Male Female Male

Anemia (Hb<11) Malaria

Figure A.2: Baseline Malaria and Anemia Prevalence, by Demographic GroupNotes: Data from Spring 2007 baseline survey. The bars represent the results of blood testing for anemia(n = 2, 532) and malaria (n = 2, 561) prevalence.

A.21

0.1.2.3.4.5

0-4

5-14

15-5

9>

=60

Age

gro

up

CM

FFr

ee

A. M

alaria

Pre

valen

ce, M

ales

0.1.2.3.4.5

0-4

5-14

15-5

9>

=60

Age

gro

up

CM

FFr

ee

B. M

alaria

Pre

valen

ce, F

emale

s0.2.4.6.81

0-4

5-14

15-5

9>

=60

Age

gro

up

CM

FFr

ee

C. A

nem

ia (H

b<11

g/dl

) Pre

valen

ce, M

ales

0.2.4.6.81

0-4

5-14

15-5

9>

=60

Age

gro

up

CM

FFr

ee

D. A

nem

ia (H

b<11

g/dl

) Pre

valen

ce, F

emale

s

Fig

ure

A.3

:P

ost-

inte

rven

tion

Mal

aria

and

Anem

iaP

reva

lence

,by

Age

and

Gen

der

Not

es:

Col

um

ns

show

anem

iaor

mal

aria

pre

vale

nce

inth

esp

ecifi

cage-

gen

der

gro

up

,by

exp

erim

enta

larm

.E

ach

colu

mn

als

od

isp

lays

95%

con

fid

ence

inte

rval

s,ro

bu

stto

intr

a-vil

lage

corr

elat

ion

.

A.22

A.23

A. Usage 90% B. Usage 45%

Coverage rate Coverage rate

Rel

ati

ve

Tra

nsm

issi

on I

nte

nsi

ty

Figure A.4: Figure 1: Protective Power of ITNs vs. Community Coverage

Source: Calculations from the epidemiological model in Killeen et al. (2007). The graphs can be pro-duced using the spreadsheet provided by Killeen et al. at http://www.ncbi.nlm.nih.gov/pmc/articles/PMC1904465/bin/pmed.0040229.sd001.xls. Coverage is defined as the fraction of individuals using an ITNeach night, while the relative transmission intensity is the proportional reduction of infectious bites for users(continuous lines) and non-users (dashed line in graph A). The label ‘usage’ refers to the fraction of time ofnormal exposure during which the individual is actually protected by the ITN.



A.24

-.4

-.2

0

.2

.4

.6

.8

0 .1 .2 .3 .4 .5(# program ITNs)/(Village pop.)

Change OLS OLS, no outliers

(A) Change in Malaria prevalence (Free)

-.4

-.2

0

.2

.4

.6

.8

0 .1 .2 .3 .4(# program ITNs)/(Village pop.)


(B) Change in Malaria prevalence (MF)

-.4

-.2

0

.2

.4

.6

.8

0 .05 .1 .15 .2 .25# BISWA members/total population


(C) Change in Malaria prevalence (Controls)

Figure A.5: Malaria Prevalence vs. Intensity of ITNs Distribution

Note: Data from spring 2007 and winter 2008-09. Each circle in the graphs represents a village. Thecontinuous lines in each graph show fitted values of a village-level OLS regression through the points. Thedashed lines show fitted values when we exclude villages with coverage larger than 0.35 (graphs A and B) orwith more than 20% BISWA membership (graph C). The point estimates of the slopes and the correspondingheteroskedasticity-robust standard errors (in parenthesis), using all villages or excluding outliers respectively,are as follows: (A) .59 (.17)*** and −.02 (.37); (B) −.54 (.32) and −.72 (.54); (C) −.76 (.43)* and −.31(.53). Statistical significance is indicated with * (10% level), ** (5%) and *** (1%).

Tab

leA

.7:

Com

par

ison

ofSam

ple

Villa

ges

vs.

Ove

rall

Villa

geP

opula

tion

inStu

dy

Dis

tric

ts

(1)

(2)

(3)

(4)

(5)

(6)

(7)

Mea

ns,

by

villa

geca

tego

ryno.

ofT

ests

(p-v

alues

)N

otin

sam

ple

Con

trol

,n

=47

Fre

e,n

=47

MF

,n

=47

Villa

ges

H0:

All

H0:

Exp

er.

equal

arm

seq

ual

Are

aof

Villa

ge(i

nhec

tare

s)27

5.2

413.

147

6.4

417.

489

910.

000*

**0.

608

Num

ber

ofH

ouse

hol

ds

121.

526

1.4

359.

028

4.3

8991

0.00

0***

0.52

6Sch

edule

dC

aste

pop

ula

tion

(%)

0.13

40.

164

0.16

40.

173

8630

0.01

2**

0.92

1Sch

edule

dT

rib

ep

opula

tion

(%)

0.47

80.

328

0.37

20.

321

8630

0.00

0***

0.59

7F

emal

es0.

501

0.49

70.

496

0.49

986

300.

128

0.76

3P

rim

ary

school

0.74

60.

936

0.97

90.

936

8991

0.00

0***

0.43

2M

iddle

school

0.23

60.

383

0.59

60.

447

8991

0.00

0***

0.09

6*Sec

ondar

ysc

hool

0.12

90.

319

0.40

40.

298

8991

0.00

0***

0.52

3H

ospit

al0.

002

0.00

00.

021

0.00

089

910.

001*

**0.

312

Num

ber

ofP

rim

ary

Hea

lth

Cen

tres

0.02

50.

106

0.06

40.

064

8991

0.13

20.

712

Num

ber

ofP

rim

ary

Hea

lth

Sub

Cen

tres

0.10

50.

170

0.23

40.

213

8991

0.02

9**

0.72

7W

ell

Wat

er0.

815

0.83

00.

872

0.80

989

910.

692

0.67

8T

ank

Wat

er0.

557

0.70

20.

723

0.74

589

910.

000*

**0.

899

Riv

erW

ater

0.12

00.

106

0.17

00.

149

8991

0.74

70.

643

Can

al0.

050

0.12

80.

149

0.12

889

910.

034*

*0.

943

Num

ber

ofP

ost

Offi

ce0.

158

0.23

40.

383

0.25

589

910.

003*

**0.

246

Num

ber

ofT

elep

hon

eco

nnec

tion

s0.

285

0.53

20.

617

0.55

389

910.

000*

**0.

682

Bus

serv

ices

0.22

80.

255

0.29

80.

298

8991

0.49

90.

866

Num

ber

ofC

omm

erci

alB

anks

0.02

70.

064

0.06

40.

085

8991

0.24

20.

906

Num

ber

ofA

gric

ult

ura

lC

redit

Soci

etie

s0.

027

0.08

50.

106

0.10

689

910.

043*

*0.

919

Appro

ach

-P

aved

Roa

d0.

332

0.38

30.

426

0.36

289

910.

506

0.81

3D

ista

nce

from

the

nea

rest

Tow

n(i

nK

ilom

eter

s)34

.934

.325

.226

.189

910.

000*

**0.

445

Ele

ctri

city

for

Dom

esti

cuse

0.46

50.

702

0.57

50.

681

8991

0.00

0***

0.38

9E

lect

rici

tyof

Agr

icult

ura

luse

0.06

60.

106

0.06

40.

149

8991

0.34

60.

386

Wet

Ric

e(i

rrig

ated

)cu

ltiv

ated

Are

a(%

)0.

075

0.15

10.

188

0.18

388

750.

000*

**0.

727

Dry

Ric

e(n

otir

riga

ted)

cult

ivat

edA

rea

(%)

0.42

20.

504

0.48

30.

510

8875

0.00

5**

0.86

4

Not

es:

Dat

afr

omth

e20

01G

over

nm

ent

ofIn

dia

Cen

sus.

Th

ep

oin

tes

tim

ate

sin

colu

mn

1in

dic

ate

mea

ns

invil

lages

not

incl

ud

edin

the

base

lin

esa

mp

le,

wh

ile

esti

mat

esin

colu

mn

s2

to4

ind

icate

mea

ns

inst

udy

vil

lages

that

bel

on

gto

the

gro

up

ind

icate

din

the

colu

mn

hea

der

.T

he

figu

res

inco

lum

n6

are

p-v

alu

esfo

rth

enu

llhyp

oth

esis

that

the

mea

nof

the

vari

ab

lein

the

row

isth

esa

me

acr

oss

all

fou

rvil

lage

gro

up

s.T

he

p-v

alu

esin

colu

mn

7ar

efo

rth

ete

stof

equ

alit

yam

ong

the

thre

eex

per

imen

tal

arm

s.S

tati

stic

al

sign

ifica

nce

isin

dic

ate

das

***

(1%

leve

l),

**

(5%

)or

*(1

0%

).A

llte

sts

are

het

eros

ked

asti

city

-rob

ust

.

A.25

A.26

Table A.8: Attrition between Pre and Post Intervention Household Surveys

Dependent variable: Dummy = 1 if household (1) (2) (3) (4)was not re-interviewed at follow-up

Constant 0.041 0.05 0.2 0.173[0.005]*** [0.013]*** [0.108]* [0.109]

Free -0.023 -0.022 -0.021[0.014] [0.014] [0.013]

Micro-loans -0.003 -0.001 0.004[0.015] [0.015] [0.015]

log(monthly expenditure/household size) 0.011 0.014[0.012] [0.011]

# household members -0.002 -0.001[0.002] [0.002]

Access to electricity 0.011 0.011[0.010] [0.010]

BISWA Debt/(Total yearly expenditure)< 0.05 -0.01 -0.021[0.016] [0.017]

BISWA Debt/(Total yearly expenditure)> 0.25 -0.006 -0.012[0.022] [0.022]

Baseline bednets per head -0.018 -0.035[0.023] [0.021]

% Members who slept under net last night -0.009 0.002[0.016] [0.017]

% Members who sleeps regularly under net 0.001 0.009[0.017] [0.017]

Household head is male 0.008 0.025[0.019] [0.017]

Household head’s age (log) -0.05 -0.053[0.019]*** [0.020]***

Household head had any schooling -0.024 -0.029[0.013]* [0.012]**

% malaria +ve in household -0.005[0.013]

% anemic (Hb< 11) in household 0.005[0.011]

Observations 1844 1844 1814 1645R-squared 0 0 0.01 0.02H0 : all coefficients = 0 (p-values) 0.11 0.21 0.14

Notes: OLS estimates. Standard errors (in brackets) are robust to intra-village correlation. Asterisks indicatesignificance at the 10 (*), 5 (**) and 1% (***) level. All regressions include observations from 141 clusters(villages). The smaller sample size in columns 3 and 4 relative to columns 1 and 2 is due to missing valuesin one or more regressors.

A.27

Table A.9: Tests of Balance for Characteristics of Cash Villages (p-values)(1) (2) (3)

Cash vs. Cash: Cash:Other Study New vs. Low vs.

villages Control High price

Area of Village (in hectares) 0.862 0.822 0.196Number of Households 0.961 0.997 0.077*SC population (%) 0.378 0.221 0.785ST population (%) 0.006 0.627 0.476Females 0.884 0.931 0.964Middle school in village 0.78 0.752 0.539Secondary school in village 0.594 0.68 0.108Primary Health Centre 0.9 0.08* 0.56Primary Health Sub Centre 0.642 0.262 0.267Well Water 0.224 0.573 1Tank Water 0.046** 0.213 1River Water 0.68 0.283 0.222Canal 0.589 0.283 0.687Post Office 0.637 0.933 0.096*Telephone connection 0.836 0.874 1Bus services 0.393 0.623 0.194Agricultural Credit Societies 0.693 0.08* 0.56Paved Road 0.063* 0.066* 0.214Distance from the nearest Town (in Kms) 0.18 0.807 0.75Electricity for Domestic use 0.119 0.378 1Electricity of Agricultural use 0.719 0.906 0.643Wet Rice (irrigated) cultivated Area (%) 0.892 0.324 0.518Dry Rice (un-irr.) cult. Area (%) 0.41 0.342 0.741

# Villages 156 40 40

Notes: all figures are p-values of tests of equality of means of the listed village-level characteristics betweenvillages in the two groups indicated in the column header. All data are from the 2001 Census of India. Alltests are heteroskedasticity-robust. Asterisks denote statistical significance at the 10(*), 5(**) or 1%(***)level. The results in column (1) use information from the 40 Cash villages and from the remaining 116villages surveyed at baseline (47 MF, 47 Free and the 22 Control villages not included in the Cash study as‘PC’ villages). The results in column 2 use information from the 25 PC villages and the 15 New villages.The same 40 villages are also used in column 3, where they are split by whether LLINs were sold at loweror higher prices (20 per group).

A.28

Table A.10: Post-intervention Malaria Biomarkers: Testing Success Rate in Baseline House-holds

(1) (2) (3) (4) (5) (6)Absent Absent Absent Refusal Refusal Refusal

Free -0.001 -0.001 -0.009 -0.01[0.018] [0.018] [0.015] [0.015]

MF 0.006 0.005 0.018 0.017[0.018] [0.019] [0.016] [0.016]

Male, 0-5 -0.212 0.017[0.020]*** [0.013]

Female, 0-5 -0.205 0.045[0.023]*** [0.017]***

Male, 5-15 -0.121 0.017[0.018]*** [0.010]*

Female, 5-15 -0.136 0.008[0.019]*** [0.010]

Female, 15-45 -0.187 0.011[0.015]*** [0.006]*

Male, > 45 -0.133 0.003[0.017]*** [0.006]

Female, > 45 -0.212 0.036[0.018]*** [0.009]***

Constant 0.194 0.193 0.32 0.057 0.054 0.043[0.007]*** [0.013]*** [0.018]*** [0.006]*** [0.011]*** [0.012]***

Observations 9589 9589 9555 9589 9589 9555R-squared 0.0000 0.0001 0.0404 0.0000 0.0023 0.0052Clusters 141 141 141 141 141 141

Free=MF=0 0.9209 0.9343 0.2303 0.2355M=F,0-5 0.7449 0.1558M=F,5-15 0.4402 0.4505M=F,Over 45 0.0000 0.0010

Notes: Data from post-intervention household survey (Winter 2008-09). Standard errors (in brackets) arerobust to intra-village correlation. Asterisks indicate significance at the 10 (*), 5 (**) and 1% (***) level.All figures are OLS estimates of a linear probability model where the dependent variable is indicated inthe column header. Both absence and refusal refer to malaria RDTs, but the figures for Hb are almostidentical. All regressions include only observations from all members (at the time of the follow-up) of the1768 households interviewed at baseline and re-contacted during the follow-up survey.

A.29

Table A.11: Results of Rapid Diagnostic Tests Validation

RDT(1) RDT(2) RDT(3)

RDT(2) 0.7873RDT(3) 0.7844 0.8760

Microscopy 0.5274 0.6131 0.5968

Microscopy-ve +ve

Tester 1 -ve 129 1RDT +ve 45 30

Microscopy-ve +ve


Microscopy-ve +ve


Notes: Data from July 2009. The results refer to tests of 205 blood samples collected from individuals withmalaria symptoms in 3 villages in Rourkela district (Orissa). The figures in the sub-table on top are samplecorrelations between the results as read by the tester indicated in the column header and the one indicated inthe row. The figures in the three sub-tables underneath indicate the details of the sample joint distributionsof the test results as read by each tester vs. microscopy. Testers 1 and 2 were part of the field team thatconducted blood tests during the follow-up household survey. Tester 3 was the most senior survey monitorin the team.

A.30

Table A.12: Knowledge of Causes of Malaria and Risk Mitigating Behavior

(1) (2) (3) (4)

Means Test of equalityControl Free MF (p-values)

(A) Causes of malaria

Drinking contaminated water 0.105 0.059 0.073 0.055**Mosquito bites 0.845 0.892 0.854 0.058*Contaminated environment 0.116 0.131 0.148 0.447Don’t know 0.037 0.025 0.051 0.065*

(B) Malaria-avoiding behavior

Nets 0.819 0.866 0.830 0.139ITNs 0.023 0.023 0.017 0.718Proper clothing (long sleeves etc.) 0.004 0.008 0.010 0.268Avoid drinking contaminated water 0.076 0.054 0.058 0.471Insecticides 0.009 0.008 0.017 0.352Repellents/mosquito coils 0.030 0.020 0.020 0.554Smoke 0.016 0.023 0.022 0.622Clearing stagnant water 0.028 0.021 0.022 0.702Cleaning drainage system/sewage 0.054 0.075 0.087 0.093*Avoiding contaminated environments 0.158 0.170 0.211 0.151Proper diet 0.051 0.039 0.037 0.618Medicine 0.042 0.033 0.066 0.058*Other ways 0.035 0.021 0.027 0.469Don’t know 0.035 0.030 0.024 0.608

(C) Residual spraying of walls

Inner walls sprayed in 2008-09 0.403 0.368 0.296 0.242Outer walls sprayed in 2008-09 0.531 0.481 0.442 0.580

(D) Number of nets from other sources in 12 months before follow-up survey (per household)

From Government/health centers 0.051 0.054 0.136 0.321From NGOs other than BISWA 0.004 0.000 0.019 0.328Purchased from the market 0.678 0.139 0.511 0.000***

Notes: Data from follow-up survey (winter 2008-09). Only panel households are included (n = 1, 768).The figures in panels A and B show proportions of respondents who list, un-prompted, the cause/behaviorindicated in the row header. The p-values in column 4 are calculated for a test of the joint null hypothesisthat means are identical across experimental arms. All tests are robust to the presence of intra-villagecorrelation of residuals. Asterisks in column 4 indicate significance at the 10 (*), 5 (**) and 1% (***) level.

Tab

leA

.13:

Sel

f-re

por

ted

Mal

aria

Indic

es:

Bas

elin

ean

dF

ollo

w-u

pD

iffer

ence

sin

Lev

els

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

(11)

(12)

(13)

(14)

Cas

esw

ith

inN

um

ber

ofD

ays

ofw

ork

Hea

lth

Exp

end

itu

res

for

Cos

tsp

aid

Cos

tsp

aid

wit

ha

mon

thof

epis

od

esor

sch

ool

lost

exp

end

itu

res

dru

gs/d

oct

ors

wit

hdeb

tlo

wer

con

sum

pti

onsu

rvey

(las

t6

mon

ths)

(las

t6

mon

ths)

(las

t6

mon

ths)

(las

t6

mon

ths)

(las

t6

mon

ths)

(las

t6

mon

ths)

Bas

elin

eF

ollo

w-u

pB

asel

ine

Fol

low

-up

Bas

elin

eF

ollo

w-u

pB

asel

ine

Fol

low

-up

Bas

elin

eF

ollo

w-u

pB

asel

ine

Fol

low

-up

Bas

elin

eF

ollo

w-u

p

Sel

f-re

por

ted

Mal

aria

Fre

ed

istr

ibu

tion

=1

0.00

1-0

.001

0.03

2-0

.011

0.71

3-1

.231

189

-510

620

0.08

2-0

.033

0.09

3-0

.023

[0.0

02]

[0.0

03]

[0.0

15]*

*[0

.019

][0

.854

][1

.086

][1

21]

[178

][6

5][1

04]

[0.0

43]*

[0.0

51]

[0.0

53]*

[0.0

21]

Mic

ro-l

oan

s=1

0.00

20.

001

0.03

1-0

.018

0.92

-1.4

7516

3-1

0611

2-7

50.

032

-0.0

790.

090.

003

[0.0

03]

[0.0

03]

[0.0

15]*

*[0

.017

][0

.928

][0

.985

][1

31]

[161

][6

7]*

[86]

[0.0

33]

[0.0

45]*

[0.0

53]*

[0.0

26]

Con

stan

t(C

ontr

ol)

0.00

80.

007

0.09

20.

115

4.26

75.

779

625

863

318

487

0.15

10.

220.

244

0.06

9[0

.002

]***

[0.0

02]*

**[0

.009

]***

[0.0

13]*

**[0

.572

]***

[0.8

33]*

**[8

6]**

*[1

19]*

**[4

3]**

*[6

3]**

*[0

.022

]***

[0.0

38]*

**[0

.031

]***

[0.0

17]*

**

Ob

serv

atio

ns

9684

9598

9684

9598

1768

1768

1768

1768

1768

1768

1768

1768

1768

1768

Fre

e=M

F=

0(p

-val

ue)

0.69

920.

9247

0.04

02**

0.57

860.

5499

0.32

180.

2523

0.75

720.

1526

0.56

230.

1557

0.18

560.

111

0.40

74F

ree=

MF

(p-v

alu

e)0.

7803

0.69

310.

949

0.68

430.

8315

0.78

060.

8438

0.55

280.

929

0.35

480.

2648

0.27

550.

9574

0.26

88

Sel

f-re

por

ted

Mal

aria

and

Fev

erca

ses

Fre

ed

istr

ibu

tion

=1

0-0

.006

0.02

1-0

.049

-0.3

34-3

.213

8-2

1713

-75

0.05

4-0

.135

0.11

1-0

.057

[0.0

06]

[0.0

07]

[0.0

21]

[0.0

40]

[1.0

55]

[1.5

40]*

*[1

41]

[255

][8

0][1

45]

[0.0

60]

[0.1

11]

[0.0

89]

[0.0

74]

Mic

ro-l

oan

s=1

0.00

30.

005

0.04

1-0

.069

1.23

6-3

.571

230

-344

143

-193

0.03

2-0

.186

0.14

60.

013

[0.0

06]

[0.0

07]

[0.0

22]*

[0.0

37]*

[1.1

74]

[1.3

69]*

*[1

62]

[238

][9

0][1

32]

[0.0

59]

[0.1

17]

[0.0

86]*

[0.0

80]

Con

stan

t(C

ontr

ol)

0.02

60.

036

0.21

50.

458

7.84

713

.125

1,10

12,

011

566

1,11

10.

315

0.69

10.

613

0.26

2

[0.0

04]*

**[0

.005

]***

[0.0

15]*

**[0

.028

]***

[0.7

88]*

**[1

.190

]***

[108

]***

[184

]***

[62]

***

[99]

***

[0.0

40]*

**[0

.088

]***

[0.0

60]*

**[0

.055

]***

Un

itof

obse

rvat

ion

9684

9598

9684

9598

1768

1768

1768

1768

1768

1768

1768

1768

1768

1768

Fre

e=M

F=

0(p

-val

ue)

0.86

110.

2735

0.17

910.

1667

0.35

910.

0327

**0.

2735

0.35

350.

211

0.33

50.

6571

0.26

930.

2152

0.60

81F

ree=

MF

(p-v

alu

e)0.

6294

0.11

370.

3815

0.57

750.

1621

0.76

360.

1434

0.58

580.

1207

0.39

130.

7137

0.61

840.

7036

0.36

25

Not

es:

Dat

afr

ombas

elin

e(S

pri

ng

2007

)an

dp

ost

-inte

rven

tion

hou

seh

old

surv

eys

(Win

ter

2008-0

9).

All

resu

lts

are

OL

Ses

tim

ate

sof

diff

eren

ce-i

n-

diff

eren

ces

mod

els.

All

outc

omes

refe

rto

mal

aria

an

dfe

ver

epis

od

esd

iagn

ose

das

such

by

the

resp

on

den

t.M

on

etary

valu

esare

in2008-0

9R

up

ees.

Reg

ress

ion

sin

colu

mn

s1-

4ar

ees

tim

ated

atth

ein

div

idu

al

level

,w

hil

ere

gre

ssio

ns

5-1

4are

esti

mate

dat

the

hou

seh

old

leve

l.S

tan

dard

erro

rs(i

nb

rack

ets)

and

test

sar

ero

bu

stto

intr

a-vil

lage

corr

elati

on

.A

ster

isks

ind

icate

sign

ifica

nce

at

the

10

(*),

5(*

*)

an

d1%

(***)

level

.

A.31

A.32

Table A.14: Impact of Intervention on Self-reported Malaria and Fever Indices

(1) (2) (3) (4) (5) (6) (7)Malaria Malaria and fever episodes in 6 months before interviewor fever Number of Days of work Health expenditures # episodes # episodes paidprevious episodes or school lost All Doctors paid for for with lowermonth & drugs with debt consumption

Free distribution= 1 -0.011 -0.08 -2.9 -225 -87 -0.189 -0.168[0.009] [0.038]** [1.42]** [251] [131] [0.101]* [0.115]

Micro-loans= 1 0.002 -0.107 -4.8 -575 -336 -0.218 -0.132[0.009] [0.036]*** [1.44]*** [214]*** [114]*** [0.105]** [0.106]

Constant (Control) 0.01 0.243 5.3 910 545 0.376 -0.351[0.006] [0.028]*** [1.03]*** [166]*** [81]*** [0.078]*** [0.075]***

Endline level (Control) 0.036 0.458 13.1 2,011 1,111 0.691 0.262

Unit of observation Individual Individual Household Household Household Household HouseholdObservations 8684 8684 1768 1768 1768 1768 1768Free=MF=0 (p-value) 0.331 0.0126** 0.0045*** 0.026** 0.0113** 0.0849* 0.2808Free=MF (p-value) 0.1653 0.4275 0.1708 0.1336 0.0588* 0.7577 0.7533

Notes: Data from baseline (Spring 2007) and post-intervention household surveys (Winter 2008-09). Allresults are OLS estimates of difference-in-differences models. All outcomes refer to malaria and fever episodesdiagnosed as such by the respondent. Monetary values are in 2008-09 Rupees. Standard errors (in brackets)and tests are robust to intra-village correlation. Asterisks indicate significance at the 10 (*), 5 (**) and 1%(***) level.

A.33

Table A.15: Malaria Prevalence and Spatial Distribution of BISWA Households within Vil-lages

(1) (2) (3) (4) (5)Radius around index household (in meters)

5 10 20 30 40

# Households within radius, αP -0.016 -0.001 0.001 0.002 -0.002[0.029] [0.030] [0.015] [0.010] [0.013]

# BISWA Households within radius, αB 0.022 0.024 0.010 0.000 0.005[0.068] [0.049] [0.053] [0.018] [0.028]

# Households within radius×Free, τP 0.028 -0.001 0.001 -0.001 0.002[0.039] [0.031] [0.015] [0.010] [0.013]

# BISWA Households within radius×Free, τB -0.053 -0.035 -0.013 0.001 -0.007[0.077] [0.054] [0.054] [0.020] [0.028]

Observations 611 611 611 611 611

Notes: Data on malaria infection from 2008-09 post-intervention survey in 11 villages (4 Control and 7 Free).The dependent variable is a dummy variable for malaria infection of the individual, measured using RDTs.Standard errors (in brackets) are calculated using block bootstrap, with 250 replications and using the villageas block. None of the coefficients in the table is significant at standard levels. All regressions include villagefixed effects.

Micro-loans, Insecticide-Treated Bednets and Malaria: Evidence …tarozzi/TarozziEtAl2013RCT.pdf · 2013-09-16 · Micro-loans, Insecticide-Treated Bednets and Malaria: Evidence from

Documents